Read spacing-readable dense images through resolved hs2p plans

[clemsgrs]

Summary

• classify each dense-image run into exactly one raster or hs2p spacing-readable reader regime, including content-based capability probing for readers such as OpenSlide
• resolve immutable per-image backend, source spacing, native level, tolerance, read geometry, and effective spacing plans in the parent before resume filtering, then round-trip them to distributed ranks
• read complete selected levels through hs2p, area-downsample without upsampling, enforce strict post-read target geometry, and persist the complete plan in sidecars and compatibility identity
• document the public reader/spacing contract and cover native, downsampled, default/override, backend, serialization, metadata, resume, mixed, unsupported, and OpenSlide-backed examples

Why

Dense extraction previously handled only Pillow raster inputs, so pre-cropped WSI-format images could not be read at a run-level physical scale or resumed against a fully resolved read identity. The parent now owns resolution and ranks consume the exact immutable plan, preventing backend/level policy drift across processes.

User impact

Callers can pass hs2p-readable pre-cropped images to Model.embed_images_dense, request an explicit physical spacing or use one unambiguous model default, override source level-0 spacing when needed, and receive sidecars that distinguish declared, source, native-read, and effective spacing. Mixed regimes, unsupported inputs, missing metadata, invalid overrides, and target-geometry mismatches fail before model inference with actionable errors.

Validation

• focused dense-image suite: 89 passed, 1 deselected before the final classifier review; final classifier regressions: 4 passed
• full suite: 757 passed, 14 skipped; the sole failure is the pre-existing tests/test_dense_image_shard.py::test_multi_rank_matches_single_rank exact CPU tensor comparison, reproduced on clean main
• docs: python -m sphinx -W -b html docs ... succeeded
• final /code-review: no actionable Standards or Spec findings

Closes #259