feat: preserve Leica confocal channel identity (dye, excitation, emission) in OME-Zarr

[aelefebv]

Problem

zarrmony's promise is "convert any bioimage to OME-Zarr, preserving metadata." For
Leica confocal .lif files that promise was silently broken: every channel's
fluorophore, excitation laser line, emission detection band, and detector were dropped.

Root cause is two gaps in bioio-lif, now filed + fixed upstream
(bioio-devs/bioio-lif#51, PR bioio-devs/bioio-lif#52):

1. It names confocal channels by the bare ChannelDescription/@LUTName — a display LUT
   color ("Blue", "Gray", "Green", …), not the fluorophore.
2. It doesn't implement ome_metadata (inherits bioio_base's NotImplementedError), so
   api._try_get_ome_image fell back to a <MetadataOnly/> stub — no OME-XML channel
   metadata at all. (CZI/ND2 plugins do implement it; LIF was the odd one out.)

The real identity is all present in the LIF XML, which we already preserve verbatim at
OME/source/raw.lif.xml. This PR recovers it.

What changed

• src/zarrmony/metadata/lif_channels.py (new) — a pure, stdlib-only,
  dependency-free extractor:
  • extract_channels(scene_xml) → per-channel {index, dye, fluor, detector, excitation_nm, emission_low_nm, emission_high_nm}. The join is the careful part: the
    real acquisition sequences are the ATLConfocalSettingDefinition children of
    LDM_Block_Sequential/LDM_Block_Sequential_List, indexed by each channel's
    SequentialSettingIndex — which correctly ignores the duplicate …Master block and
    a phantom 620 nm setting that would otherwise mislabel channels. Excitation is the
    active laser line paired to the active detector by spectral order; emission is the
    detector's MultiBand window.
  • channels_to_ome_channels(...) / channels_to_omero(...) — project that identity into
    OME-XML <Channel> (Name/Fluor/ExcitationWavelength/EmissionWavelength) and omero
    Dye (NNN nm) display labels + colors.
  • Hardened: rejects DTDs/entity expansion (billion-laughs/XXE), caps input size, and
    fails closed (returns [] / falls back) on any malformed input — never raises,
    never hangs.
• src/zarrmony/api.py — for LIF the per-scene path now builds a real OME Image
  (with channels) and meaningful omero labels instead of the stub. LIF-only (keyed on
  reader.scene_root), fail-safe (any extraction error reverts to the prior
  name-based path — metadata never breaks a conversion), and channel-count vs SizeC
  guarded so it can't mislabel. CZI/ND2/non-LIF paths are untouched.
• Tests — a committed pruned LIF scene fixture (tests/fixtures/lif_confocal_7ch.xml,
  no multi-GB file needed) + extractor, projection, and api-wiring (fail-safe) suites,
  including the entity-bomb / non-plate-scene_root-raises / count-mismatch edges.
  The same extraction has been ported upstream into bioio-lif's own ome_metadata
  (Confocal channel identity: implement ome_metadata and name channels by fluorophore bioio-devs/bioio-lif#52); once that ships, this LIF-specific layer can be retired in
  favor of it. Until then, zarrmony carries it so conversions preserve the metadata today.

Verification

On a real 7-channel Leica STELLARIS confocal scene (CohortAI37), the converted store now
carries, in order: DAPI (405 nm), ALEXA 594 (590 nm), ALEXA 750 (753 nm),
ALEXA 488 (499 nm), ALEXA 647-R-PE (653 nm), ALEXA 555 (553 nm), ALEXA 700 (696 nm)
— matching the dyes/lasers/emission bands cross-checked against the acquisition metadata.
uv run ruff check / pytest / black --check . / npx prettier --check . all green.

Compatibility / scope

Additive and back-compatible: no change to CZI/ND2/non-LIF behavior, the audit schema, or
existing outputs. Out of scope (deferred): emission-derived omero colors; the bf2raw/plate
LIF paths (per-scene is the default LIF layout); widefield channel enrichment.

[aelefebv]

Note: collapse the in-tree extractor onto bioio-lif once #52 lands

This PR deliberately carries an in-tree, temporary copy of the confocal channel extractor (src/zarrmony/metadata/lif_channels.py) plus bespoke per-scene wiring in api.py (_lif_scene_channels / _lif_ome_image). It exists only because released bioio-lif (1.4.2) does not implement Reader.ome_metadata — it raises NotImplementedError — so zarrmony has to extract channel identity itself for now.

The proper home is bioio-devs/bioio-lif#52 ("implement ome_metadata and name channels by fluorophore"), which contains the same extractor and locates each scene's XML correctly via self.metadata.findall(".//Image")[self.current_scene_index].

Once #52 is merged and released, this PR's interim layer should be removed:

1. Bump the bioio-lif minimum pin to the release that includes feat(lif): extract objective lens metadata (magnification, NA, immersion) into audit + OME-XML #52.
2. Delete the in-tree extractor metadata/lif_channels.py and the bespoke wiring (_lif_scene_channels, _lif_ome_image) in api.py.
3. Build the per-scene OME Image from reader.ome_metadata directly, keeping the existing fail-safe fallback and the channel-count-vs-SizeC guard.
4. Drop the tests that exercise the in-tree copy; keep one integration test asserting channels still surface via the delegated path.

Interim correctness fix (this branch): the wiring originally located the scene via reader.scene_root, which is bioio-lif's plate row/column locator and raises ValueError on ordinary confocal scenes — silently disabling the whole feature for the confocal .lif files it targets (output fell back to LUT-color labels + a MetadataOnly stub). This branch fixes that by aligning the scene-location with #52's .//Image[current_scene_index] approach, so the two implementations stay consistent until the delete above.