Stop copying train_mod2 into novel's model artifact

[rcannood]

Describe your changes

novel_train writes a full copy of the training target into its model directory:

adata = input_train_mod2.copy()
...
adata.uns["model_dim"] = {"mod1": n_vars_mod1, "mod2": n_vars_mod2}
adata.write_h5ad(f"{par['output']}/train_mod2.h5ad", compression="gzip")

but novel_predict only ever touches uns on it:

mod2        = input_train_mod2.uns['modality']
n_vars_mod1 = input_train_mod2.uns["model_dim"]["mod1"]
n_vars_mod2 = input_train_mod2.uns["model_dim"]["mod2"]
rem_var     = input_train_mod2.uns["removed_vars"]

so the matrix, obs and var are written, published and re-read for nothing. Keeping just the uns:

modality:  ATAC
model_dim: {'mod1': 256, 'mod2': 1500}
removed_vars via .get(): None

artifact size old: 0.69 MB   (test fixture, 171 x 1500)
artifact size new: 0.53 MB

The fixture is small enough that h5ad overhead dominates; on the real datasets this is the difference between carrying a full expression matrix per model and not.

I did consider having novel_predict read --input_train_mod2 directly -- the wrapper already passes it -- but that argument is optional in comp_method_predict.yaml, so the model artifact staying self-describing seemed safer.

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!