Fix output dimension in simple_mlp_predict

[rcannood]

Describe your changes

simple_mlp_predict sizes the MLP output layer with input_test_mod1.n_vars -- the source modality's width. But training sizes it with y.shape[1], i.e. train_mod2.n_vars:

net = MLP(X.shape[1], y.shape[1], ymean, config)   # simple_mlp_train

So MLP.load_from_checkpoint(..., out_dim=input_test_mod1.n_vars, ...) raises a state-dict size mismatch on fc3 whenever n_vars(mod1) != n_vars(mod2) -- which is every real dataset (GEX 13953 vs ADT 134). The GEX2ATAC branch has the same mixup: ymean has length n_vars(mod2) but is broadcast against np.ones([n_obs, input_test_mod1.n_vars]).

input_train_mod2 is already read in on line 49, so it's a two-line change.

The reason this never showed up in CI is worth a note: in resources_test the two modalities happen to have identical widths -- 134x134 for bmmc_cite, 1500x1500 for bmmc_multiome -- so viash test can't distinguish n_vars(mod1) from n_vars(mod2) and passes. I'll follow up with a fixture that has unequal widths so this class of bug can't hide again.

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!