Fix data section loss, chain indent idempotency, and else comments

[sorafujitani]

Three formatter bugs found by ad-hoc property testing over Rails-style code (crash / idempotency / AST equivalence / comment preservation, plus a DATA-section property the existing corpus check does not cover).

What was broken

The `END` data section was silently dropped. `Rfmt.format("puts DATA.read\n\n__END__\nhello\n")` returned only `"puts DATA.read\n"`, destroying any script that reads `DATA`. The section is not part of Prism's AST, and the formatter rebuilds output from the AST, so nothing ever emitted it. Fix: `NativeAdapter::parse` records `data_loc()`'s start offset in the root metadata, and `Formatter::format` re-appends the verbatim source slice after printing. `data_loc` comes from Prism, so a fake `END` inside a heredoc cannot false-positive.

Chain continuation indentation was not idempotent. The chain reformatter baked the statement's input column into a `Doc::Text`, but the doc engine places the statement at its output position. On misindented input (conflict resolutions, generated code) the first pass under- or over-indented continuations and the second pass moved them again. Fix: `reformat_chain_lines` became `reformat_chain_doc`, emitting continuation lines as `hardline` docs inside `indent(...)` so the printer anchors them at render time. Multi-line arguments keep their offset relative to the chain; heredoc bodies are emitted through `literalline` verbatim, because for plain/dash heredocs the leading whitespace is string content (a lexical heredoc scanner marks those lines; ambiguity fails toward verbatim, the semantics-safe direction).

A trailing comment on `else` moved below the line, where it reads as annotating the branch body. The `ElseNode` arm never consumed the else-line comment as a trailing comment, so the body's leading-comment path claimed it. Fix mirrors the existing if-header and end-line handling.

Verification

• Failing repro specs land first in history (`spec/end_data_section_spec.rb`, `spec/chain_indent_idempotency_spec.rb`, `spec/else_trailing_comment_spec.rb`), fixes on top.
• `bundle exec rspec`: 177 examples, 0 failures. `cargo test`: 139 + 5 + 3 passed. `scripts/corpus_check.rb`: 49/49.
• A 77-case Rails-idiom corpus (migrations, models, controllers, routes, jobs, RSpec, Rake DSL, mangled-indentation variants, lexical traps) passes all properties including the new DATA-section check.

Known follow-ups (pre-existing, not regressions)

• A variable-assigned chain ending in a `do…end` block under-indents the block body on misindented input (output is stable and semantically equivalent).
• `case/when` and `begin/rescue` `else` lines still relocate their trailing comment; the fix here covers `if`/`unless`/`elsif` only.
• Two lexical heredoc detectors now coexist (`heredoc_body_lines` and `statement_contains_heredoc_tail` in if_unless.rs); the older, cruder one could be reimplemented on the new scanner.