Make the test fixture non-square

[rcannood]

Describe your changes

Needs an aws s3 sync to take effect -- see below. Happy to hold this until the rest of the series has landed.

Three of the bugs in this review series were invisible to CI for the same reason: in resources_test the two modalities have identical widths.

bmmc_cite      134 x 134
bmmc_multiome 1500 x 1500

No real dataset looks like that -- bmmc_cite is 13953 GEX features against 134 ADT proteins. So a component can use n_vars(mod1) where it means n_vars(mod2) and viash test goes green, then it dies on the first full run. That's exactly what #24 was.

Adding a trim_mod1_features.py step to test_resources.sh that drops a quarter of mod1's features after process_datasets and before the models are pre-trained (the ordering matters -- the stored novel and simple_mlp checkpoints are sized from mod1, so they have to be trained against the trimmed version).

widths after trimming:
  bmmc_cite/normal       train   100 x 134    test   100 x 134    distinguishable
  bmmc_cite/swap         train   100 x 134    test   100 x 134    distinguishable
  bmmc_multiome/normal   train  1125 x 1500   test  1125 x 1500   distinguishable
  bmmc_multiome/swap     train  1125 x 1500   test  1125 x 1500   distinguishable

Trimmed files still satisfy file_train_mod1.yaml -- subsetting on var keeps every required slot:

  layers['counts']         ok
  layers['normalized']     ok
  obs['batch']             ok
  var['hvg']               ok
  var['hvg_score']         ok
  uns['dataset_id']        ok
  uns['normalization_id']  ok

I did look at fixing this in the common fixture instead, which would be the more principled place -- but resources_test/common/openproblems_neurips2021 is shared with other tasks, so making it asymmetric for our benefit isn't ours to do. Doing it task-side keeps the blast radius here.

To take effect

This PR only changes the generation script; the resources on S3 are still square. Regenerating and syncing writes to the shared bucket, so I've left that to you:

bash scripts/create_datasets/test_resources.sh   # regenerates + prints a dryrun sync
aws s3 sync --profile op \
  resources_test/task_predict_modality \
  s3://openproblems-data/resources_test/task_predict_modality \
  --delete

(the script's own sync still has --dryrun on it, deliberately.)

Part of a series of PRs coming out of a pre-run review of the benchmark.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!