Build Okhsl and OkLCh planes three times faster, and share the dialog's work

[smelfungus]

A hue drag rebuilds the plane at every step, and on desktop an Okhsl or OkLCh plane took over 30 ms to build, so the plane trailed the thumb; a phone is several times slower. Every change here keeps results bit for bit: the 30 screenshots validate unchanged, and new tests hold the colors to what they were.

Plane at hue 200 (PlaneTimingTest, desktop)	Before	After
Okhsl S × L, 256 × 256	30.5 ms	9.9 ms
OkLCh C × L, 256 × 256	32.0 ms	10.3 ms
LCH C × L, 256 × 256	35.0 ms	33.5 ms
Okhsv S × V, 64 × 32	0.9 ms	0.4 ms

• The gamut's cusp and edges are remembered between colors (color). A plane at one hue solved the same cusp for every color and the same edge along every row. Each thread now keeps its last cusp and last sixteen outer edges, keyed by the exact bits asked with, and overwritten in place, so bulk conversions still allocate nothing per color (BulkAllocationTest). RememberedGamutTest fails if a key leaves out the gamut or the lightness. LCH gains little: its constant-hue lines curve through Oklab and never meet the same edge twice.
• Colors outside sRGB are found before the curve (color). The mapper stops in linear light for the sRGB curve and encodes only what may be inside; within a trillionth of the cube's faces the curve still decides, because encoding 1 gives 0.9999999999999999. colorsOnTheCubesSurfaceMapAsToGamutMapsThem fails without that band.
• Plane rows are packed into pixels as they are converted. A 256 × 256 build allocates its 256 KB of pixels and a 6 KB row rather than another 1.5 MB of Doubles.
• Planes share rasters. One being built is awaited rather than built again, and the last three are kept, so the side-by-side dialog's measured copy no longer builds every raster a second time, and switching back to the space just left builds nothing.
• The dialog's measured copy builds no tracks. While the dialog is side by side, a BasicChannelSlider in its stacked copy hands its track a placeholder instead of rebuilding a gradient no one sees.

The color module gains a jvmAndAndroid source set for the memo's ThreadLocal, as foundation has for its number formatter, and one-line actuals for iOS and wasm. No public API changes; all targets' tests, apiCheck, the samples and the iOS compile pass. Not measured on a phone.