…de1vec into codex/various-fixes

Read tiles directly from WSI during embedding via cucim's batched
read_region, eliminating the tar creation step. Super tiles (8x8, 4x4,
2x2 blocks) reduce the number of read calls. A custom batch sampler
(adaptive_batching) optionally aligns DataLoader batches to super tile
boundaries to avoid redundant reads.

New config options under tiling:
- on_the_fly (default true): read from WSI instead of tar
- gpu_decode (default false): experimental GPU JPEG decoding
- adaptive_batching (default false): vary batch size to match super tiles

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- remove dead try/except around gpu_decode dict assignment
- remove unused variable in read_batch
- update docstring to mention 2x2 blocks
- fix on_the_fly default mismatch in from_config fallback (False → True)
- use np.isin instead of set for DDP tile filtering

Co-Authored-By: Claude Opus 4.6 <[email protected]>

The flat arrays (tile_to_st, tile_crop_x, tile_crop_y) are the only
lookup structures used in the hot path. The per-super-tile duplicates
were never read.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

PyTurboJPEG>=2 requires libjpeg-turbo 3.0+. Ubuntu 22.04 ships 2.x via
libturbojpeg0-dev. Build libjpeg-turbo 3.1.0 from source in the build
stage and copy the shared libs to the runtime stage.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Pulls in hs2p[cucim] (cucim-cu12, cupy-cuda12x, nvidia-nvimgcodec-cu12)
and PyTurboJPEG.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

_WSDTarReadPlan now carries tile_indices directly, so _build_supertile_index
no longer needs to reconstruct member coords via coord_to_index — replaced
with iter(plan.tile_indices). Also bumps hs2p requirement to >=2.4.1 which
introduced the top-down super tile grouping and the tile_indices field.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

Threads hs2p's jpeg_backend parameter (turbojpeg/pil) through
PreprocessingConfig → _tile_slides so callers can pick the JPEG
encoder used during tar extraction.  Defaults to "turbojpeg".

Sets jpeg_backend="pil" in test_output_consistency so the tar
path matches the PIL-encoded ground truth fixtures.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

Adds num_cucim_workers (default 4) to PreprocessingConfig. In the
on-the-fly path, DataLoader num_workers is auto-derived as
cpu_count // num_cucim_workers instead of reusing speed.num_workers,
avoiding oversubscription when both are large. A log message is emitted
when the computed value differs from speed.num_workers.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

Adds num_preprocessing_workers to ExecutionOptions, read from
speed.num_preprocessing_workers (fallback speed.num_workers).
_tile_slides now uses num_preprocessing_workers for hs2p, while
execution.num_workers (speed.num_dataloader_workers) remains the
DataLoader knob for the tar embedding path.

Default yaml updated with explicit num_preprocessing_workers and
num_dataloader_workers keys. Backward-compatible via num_workers
fallback in from_config.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

…eading paths

The batch preprocessor was applying a bilinear resize to target_tile_size_px
before model transforms, while the old TileDataset path used PIL Lanczos and
the model's own Resize transform. For the test WSI (read_tile_size=444px,
target=224px) this produced noticeably different pixel values.

Fix: defer resizing to the model's own Resize transform when one exists,
extracting its interpolation mode (e.g. BICUBIC for Virchow) into
BatchTransformSpec.resize_interpolation and applying it via F.interpolate.
Only fall back to a bilinear pre-resize when the model has no Resize step.

The CI consistency test is updated to always run on_the_fly=True (direct WSI
reads, no JPEG round-trip) and drops jpeg_backend which is tar-path-only.
GT embedding fixture regenerated with the new consistent preprocessing.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>