Fix ICC profile tagging across export formats

[linkmodo]

Summary

An audit of every export format × colour space combination (encoding synthetic buffers through _encode_export and inspecting the resulting bytes) found four colour-profile defects. The mainstream paths — JPEG/PNG/TIFF in sRGB / Adobe RGB / ProPhoto / P3 D65 / Rec 2020, plus custom output ICC overrides — were already correct: right profile embedded, pixels converted. These were the broken cases:

1. Greyscale JPEG/WebP exported untagged, untransformed pixels. The L-mode image was handed to a CMS transform whose source profile is the RGB working-space profile; lcms refuses to build an RGB→gray transform ("cannot build transform" on every greyscale JPEG/WebP export), the exception was swallowed, and the file shipped working-TRC luma with no embedded profile — visibly inconsistent with the greyscale TIFF/PNG of the same edit. Both formats now run the same 16-bit grey re-encode as TIFF/PNG (new _greyscale_to_pil_u8 helper) and embed the grey profile.

2. Greyscale encode used a pure 1/2.2 power, but the tag carries the sRGB TRC. The bundled GrayGamma2.2.icc actually holds the sRGB TRC despite its name — the existing _JXL_COLOR comment documents this, and I verified it by probing the profile through lcms (L* at input 8/255 is 2.35, matching sRGB's 2.19, not pure-2.2's 0.44). Pixels encoded 2.2 but tagged sRGB-TRC read wrong in the shadows. The re-encode now uses the sRGB OETF, which also makes JXL's GRAY/SRGB enumerated tagging consistent with its pixels.

3. ACES/XYZ export targets silently produced untagged Adobe RGB. Neither space has a mapped/bundled ICC profile, so CMS logged "ICC profile not found", skipped the conversion, and wrote untagged working-space pixels. Two-part fix: the encoder now falls back to exporting the working space with its profile embedded (warning logged) so stale presets/DB configs can never produce a mistagged file, and the export colour-space combo no longer offers unmappable spaces (new EXPORT_COLOR_SPACES; ACES/XYZ remain available where they belong, as rawpy decode spaces).

4. Contact-sheet JPEGs were untagged. Tiles are converted to sRGB display space before compositing; the sheet JPEG now embeds the bundled sRGB profile.

Test plan

• New tests/test_export_icc.py (6 tests): greyscale JPEG tagged + matches greyscale TIFF tonally, greyscale WebP tagged, shadow-probe proving the sRGB TRC (and rejecting the old 2.2 curve), ACES fallback produces working-space pixels with the exact working-space profile bytes, EXPORT_COLOR_SPACES excludes unmappable spaces, contact-sheet sRGB bytes resolve. All four behavioral tests verified to fail against the unfixed encoder.
• Full re-run of the format × space audit: every combination now embeds a correct profile; no CMS errors.
• Existing export suites pass (test_export_color_management, test_export_jxl, test_export_webp, test_export_settings_form, test_contact_sheet, test_icc_lut, …) — 122 passed; the only failure is the known pre-existing Windows path-separator issue in test_output_dir_subfolder_of_source, unrelated.
• ruff check + ruff format clean.