Declare cell_type, and split by batch when is_train is missing

[rcannood]

Describe your changes

process_dataset reads two obs columns that the API never declares:

is_train <- which(ad1$obs[["is_train"]] == "train")
is_test  <- which(!ad1$obs[["is_train"]] == "train")
...
ct <- as.character(ad1$obs[["cell_type"]][is_test])

process_datasets/main.nf validates its inputs against file_common_dataset_mod{1,2}.yaml and drops anything that fails -- so a dataset that honours the published contract but happens to lack is_train sails through QC. In R, which(NULL == "train") is integer(0), so both index sets come back empty and we write four zero-row objects without complaining.

Reproduced on bmmc_cite with obs["is_train"] removed:

Creating train/test split
Create train objects
Create test objects
Saving output files as h5ad
exit: 0
train_mod1 -> (0, 134)
train_mod2 -> (0, 134)
test_mod1  -> (0, 134)
test_mod2  -> (0, 134)

is_train is optional

I originally declared it required, but that's wrong: obs["is_train"] is the NeurIPS 2021 competition's own split, and there's no reason another dataset would carry it. Declared optional, with a fallback that holds out a quarter of the batches -- keeping the spirit of the competition split, where test cells come from donors the method hasn't seen -- or a quarter of the cells when there's only one batch. It's seeded off --seed, so it's reproducible.

All three paths, on bmmc_cite:

=== with_is_train
Creating train/test split
Train cells: 431, test cells: 163
=== no_is_train
No obs['is_train'], holding out batches: s1d3, s4d8, s3d7
Train cells: 446, test cells: 148
=== one_batch
No obs['is_train'] and only one batch, holding out a quarter of the cells
Train cells: 446, test cells: 148

(the two fallbacks landing on 148 is a coincidence of floor(594/4); they select different cells, and no path leaks a cell between train and test.)

cell_type is required

Declared and enforced up front, so a missing annotation fails immediately rather than a few hundred lines later inside the subsampling branch.

Also added a line logging the split sizes -- "how many cells ended up on each side" was previously invisible in the logs.

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!