Add GPFM and GenBio-PathFM as supported tile encoders

[clemsgrs]

Adds two open-weight histopathology tile encoders in a single PR (one shared docker-test run).

Closes #210
Closes #211

GPFM (gpfm)

majiabo/GPFM (MIT) — ViT-L/14 DINOv2. Weights ship as a standalone GPFM.pth rather than a timm hf-hub: config, so the timm arch is built unpretrained and the checkpoint is load_state_dict-ed (MOOZY pattern). A defensive unwrapper handles wrapped/prefixed re-exports but is a no-op on the published bare state_dict.

• Verified on real weights: GPFM.pth is a bare DINOv2 state_dict, loads strict=True with 0 missing / 0 unexpected keys, pooled forward returns (1, 1024).
• Metadata: level tile, dim 1024, input_size 224, patch_size 14, supported_spacing_um=0.5, precision fp32, source majiabo/GPFM. Dense-capable (in both dense coverage lists).
• Spacing is 0.5 µm/px (20x-equivalent): upstream recommends 512px@0.25µm resized to 224, ≈ 0.5 µm/px effective, matching UNI — not the model's native acquisition magnification.

GenBio-PathFM (genbio-pathfm)

genbio-ai/genbio-pathfm — 1.1B-param custom HF AutoModel (trust_remote_code), registered as a custom TileEncoder (not timm) with non-ImageNet normalization (mean=(0.697,0.575,0.728), std=(0.188,0.240,0.187)).

• Output dim verified on real weights: the backbone is a single-channel ViT (in_chans=1, embed_dim=1536); its canonical forward splits the RGB tile into 3 single-channel images and concatenates the per-channel CLS tokens, returning (1, 4608) (= embed_dim*3, matching the card's advertised feature dim). encode_tiles returns this default forward output.
• Dense (2D-grid) support: via the model's forward_with_patches (DINOv2-style x_norm_patchtokens), which returns the three per-channel patch grids concatenated along the feature dim — folded into (B, 4608, 14, 14) at 224 (patch size 16, prefix tokens already stripped). So d (4608) equals the pooled dim. Added to the dense coverage list.
• Attention-map extraction is not supported (deliberate opt-out): the backbone uses a fused scaled_dot_product_attention (no materialized weights, no output_attentions) and encodes the three colour channels independently, so there is no single coherent CLS-over-patches attention to recover — only three separate grayscale-channel attentions. Documented in the encoder docstring and the docs attention section.
• Pooled transform aligned with the model card: Resize((224, 224)) (exact, tuple form) + custom Normalize, no CenterCrop (the card has none, and it is a no-op after a square resize). get_dense_transform is normalization-only.
• Metadata: level tile, dim 4608, input_size 224, patch_size 16, supported_spacing_um=0.5 (20x), precision fp32, source genbio-ai/genbio-pathfm. Custom GenBio AI Community License noted in the docstring.

Tests

• Metadata-contract additions: test_encoder_registry.py (EXPECTED_TILE_ENCODERS + focused test_gpfm_metadata_contract / test_genbio_pathfm_metadata_contract), test_patch_size_metadata.py (gpfm (14,14), genbio-pathfm (16,16); both dense-capable, NON_DENSE_TILE_ENCODERS now empty), test_regression_core.py (list_models("tile")).
• tests/test_gpfm_genbio_heavy.py (@pytest.mark.heavy) asserts the real-weight forward shapes and dense grid shapes — GPFM (1, 1024, 16, 16), GenBio (1, 4608, 14, 14); skip cleanly if weights are unavailable.
• Default suite stays network-free / weight-free.

Docs

README tile count 18 → 20; docs/models.rst tile-table rows for both; both added to the dense-grids coverage list (with a per-channel-concat note for GenBio), and GenBio flagged as not covered in the dense-attention section. Reference cells kept minimal (no license / normalization notes). Also dropped the stale "gated weights, requires HF_TOKEN" note from the mSTAR row.

Notes on divergence from the original plan

• GPFM spacing landed at 0.5 (effective input spacing after resize), not the initially proposed 0.25 (which conflated training magnification with effective input spacing).
• GenBio gained a dense (2D-grid) path; the original plan scoped it pooled-only. Attention grids were investigated against the real model source and deliberately left unsupported (reasoning above).