Migrate to the PowerAnalytics 1.0 metrics API

[PabloBotin]

Closes #144.

PowerAnalytics deprecated its pre-1.0 accessors (get_generation_data, get_load_data,
get_service_data, categorize_data, PowerData, …) in favor of the 1.0
Metric/ComponentSelector API — see the
"Old PowerAnalytics" notice
and tracking issue
PowerAnalytics.jl#28.
This moves PowerGraphics' internals across without breaking any public signature.

Why this is bigger than #144

Proving the migration correct meant comparing old and new API output per category, per
timestep — which is how the bugs below surfaced. That comparison has to hold on both
backends, and asserting it revealed the two ext/ recipes had drifted: each derived its own
defaults for fill, line width, draw order, title handling, empty input, save paths and
palette. Eight behaviors resolved twice, independently — the same drift that let the
bar-stacking fix in #140 land in one recipe and silently not the other.

Once those are resolved once in src/call_plots.jl and the recipes only draw, the backend is
purely a value, which src/backends.jl already modeled. The ten _plotly-suffixed functions
were then duplicating the public API without buying any dispatch, so they became a backend
key word.

This does not split into two PRs cleanly — the demand sign fix, the start_time/len fix and
the fuel-categorization rewrite live in the same functions as the backend unification, and the
migration's regression tests are written through the backend-parity harness. The commits are
ordered so the migration can still be read on its own: the first seven are the migration, the
last two the backend key word and review fixes.

1. Migration to the metrics API

plot_demand, plot_fuel and plot_results are all on the new API. plot_results no longer
constructs a PA.PowerData, and plot_powerdata(::PowerData) is deprecated to a shim.

One new export, get_demand_data, returning the demand data plot_demand draws with its
DateTime axis. Reading a single load metric does not give the same answer (see the sign bug
below) and there was no public route to those numbers — the report template needed one, and
every other public function returns a plot object.

2. backend is now a key word

plot_fuel(res)                                   # CairoMakie (default)
plot_fuel(res; backend = PlotlyLightBackend())   # PlotlyLight

The _plotly names still work but warn and forward. Passing both a _plotly name and a
backend key word raises an ArgumentError rather than letting one silently win.

Behavior users will notice:

• Both backends now select from the whole load_palette palette, so more series get a
  distinct color before the cycle repeats. PlotlyLight's default colors change.
• The default PlotlyLight save format is now html. A shared hardcoded "png" made every
  default-path save trip the extension-rewrite warning. An explicit format still wins.
• CairoMakie non-stacked draw order now matches PlotlyLight.
• WeaveExt = ["PlotlyLight", "Weave"] → "Weave". Neither the extension nor the report
  template touches PlotlyLight, so the old trigger withheld report from CairoMakie-only
  users.

docs/src/explanation/backend_parity.md records what the backends guarantee to render
identically and where they deliberately differ; test/test_backend_parity.jl enforces it.

Bugs fixed

• Demand had the wrong sign and magnitude under controllable load formulations.
  calc_load_forecast applies an unconditional -1, but PowerSimulations stores the load
  parameter with a formulation-dependent sign, so a mixed system partially cancelled —
  measured −3604.307 against a true demand of +8853.873. _demand_data(::IS.Results) now
  resolves (calc_active_power, calc_load_forecast) per concrete load type.
• The report template's Load table read calc_system_load_forecast directly, reporting a
  different number than plot_demand for the same results — the bug above, shipped in the
  template. It now goes through get_demand_data.
• The plot_fuel net-load line now includes storage charging / source input, as its comment
  always claimed; the offset was computed and dropped.
• plot_results(...; combine_categories = false) crashed with a MethodError. The docstrings
  also claimed false was the default when the code defaults to true — the documented
  default was the one that crashed.
• plot_demand(sys; start_time, len) ignored those key words on the System path while the
  docstring advertised them.
• plot_demand(...; save = dir) wrote the figure twice under two different names.
• save_plot(p, "out.HTML") silently rewrote the path on PlotlyLight, which compared against
  ".html" exactly where CairoMakie lowercases first.
• Two fuel-enum typos in the test mapping yaml (AG_BYPRODUCT, WOOD_WASTE_SOLIDS) that the
  old parser silently never matched.

Behavior notes

• Fuel categorization keeps first-match-wins semantics (most specific type, then prime mover,
  then fuel), so no component is double-counted. Specificity comes from
  PA.parse_fuel_category, not from parsing display names.
• Time windowing is applied locally by row slicing, because PowerAnalytics.compute
  mishandles window key words on simulation results.
• ext_category entries in custom mapping yamls are ignored by the new selector parser, with
  a @warn.
• The report template's Services table is gone — there is no new-API service metric.
• Unmapped generators are reported in one aggregate @error instead of one per component.

Intentionally still on the old API

plot_demand(::PSY.System) (get_load_data(::System) has no new-API equivalent),
no_datetime on user-supplied DataFrames, and the deprecated plot_powerdata(::PowerData)
shims. PowerAnalytics still maintains the old API, so these keep working until a future
breaking release.

Testing

391 pass / 0 fail, with roughly 1,600 added lines under test/.

The migration is pinned by a numeric equivalence test against the still-exported old API,
covering every category across both UC and ED problems — storage In/Out, Curtailment,
Unserved Energy, Over Generation included — agreeing bit-identically
(maxabsdiff = 0.0), not merely within tolerance.

New test files: test_backend_parity.jl (the parity contract, both backends compared
directly), test_demand_semantics.jl (the sign bug and per-formulation metric resolution),
test_fuel_categories.jl (rule specificity and enum validation),
test_fuel_stack_behavior.jl (stack composition and the net-load line), and
plot_introspection.jl (backend-agnostic helpers so parity assertions are written once).

The public API was also exercised end-to-end outside the harness against real UC and ED
results — backend selection, the deprecated names, save formats and paths, palette assignment
and demand sign — with figures rendered on both backends for inspection.

Upstream issues

Workarounds are marked # TODO upstream and filed against PowerAnalytics: broken compute
time-window key words, stale get_subselectors export, inconsistent missing-result error
types, missing calc_system_slack_down/forecast metrics, and parse_generator_categories
returning nothing. The calc_load_forecast sign bug is still to be filed.