ML4H merge: tranche 1

[will-pang]

A first pass at consolidating all the work we did so that we can merge to main. I'll probably need everyone's help in reviewing this (@joshuasteier, @Rian354, @joshuxchn, as well as @jhnwu3), but in short the updates here add in unified embedding support and tested across the different backbones (RNN, transformer, bottleneck
transformer, EHRMamba, JambaEHR) that we're interested in presenting results. In terms of modality, this focuses only on labs, labs + notes, labs + notes + cxr.

Three big updates made

• Removed all sample re-weighting work
• Removed observation windows
• Removed ICD-10 codes

What to remove

• All references in pyhealth/scripts/ which I've used to submit to condor and can be removed.

[Rian354]

Pushed four proven repairs onto this branch (fast-forward from 2b7b2c5, no rebase).

1. 9782aca Stop emitting fake missing-event placeholders ([MISSING_TEXT] / pad visits / black frames). Empty modalities are zero events.
2. 869ac8e Keep frozen text encoders in eval when Trainer calls train(), so BERT dropout cannot fire on a frozen encoder.
3. 11beefc Cache frozen [CLS] embeddings keyed on real tokens, not padded rows (padded-row keys never hit under shuffle).
4. 86fe73b Restore BaseDataset._scan_table / _scan_parquet. MEDS still calls _scan_parquet after a0f1422 deleted them. Proof: tests/test_p0_parquet_scan.py reads a real two-row parquet file.

Still local, not on this PR yet: pad-mask / fp16 fill, nested padding_idx=0, sunlab resized_images + cache write.

[Rian354]

Pushed four more proven repairs (fast-forward 86fe73b → b6fb31e). Each commit has a unit test whose docstring records the measured numbers and the pytest repro.

1. 965a87b Fill attention masks with finfo(dtype).min and use fused SDPA. -1e9 is outside fp16, so AMP overflowed on padded positions. Fused vs explicit max abs diff: 7.153e-07 (no padding), 4.768e-07 (with padding). A10 notes_labs transformer 128/2/4, 2 epochs, batch 8: bf16 5,275 s vs fp32 10,198 s (epoch-1 2.84×, mean 2.41×, VRAM -24.5%, final train loss within 0.63%). Proof: tests/test_p1_fp16_attention.py.
2. 7c4c056 Record {field}__pad_mask and skip those slots in the unified sequence. The collator padded with 0.0 and nothing marked it, so padded slots looked like real measurements at admission time. Token-budget notes crashed tensor a (7) must match b (14); reusing event pad_mask as a BERT token mask crashed shape '[96, 512]'. RNN packed lengths now clamp at 1. Proof: tests/test_p1_pad_masks.py.
3. ca9b63d Keep padding_idx=0 on nested code embeddings. Pad row is zeros and receives zero gradient. Proof: tests/test_p2_nested_padding_idx.py.
4. b6fb31e Accept resized_images and write sunlab CXR metadata to cache. Complete resized cohort 377,110 / 377,110. Hardcoded images/ was FileNotFoundError; default CXR config was KeyError: 'studytime_normalized'; read-only root now writes the CSV under cache_dir. After the layout worked (one seed, 6 epochs, 18,542 / 2,285, prevalence 0.0565): cxr_only PR-AUC 0.0602, +labs 0.3082, +notes+labs 0.4096. Those three compare with each other, not with the primary notes_labs table. Proof: tests/test_p2_sunlab_cache.py.

Repro for all four:

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTHONPATH=. python -m pytest \
  tests/test_p1_fp16_attention.py tests/test_p1_pad_masks.py \
  tests/test_p2_nested_padding_idx.py tests/test_p2_sunlab_cache.py -q

12 passed on this checkout.