Add sciPENN method (GEX->ADT)

[benjaminfreyuu]

Wraps sciPENN 0.9.6 (Lakkis et al. 2022) for CITE-seq protein prediction and registers it in run_benchmark. Unsupported directions raise.

Works around upstream bugs that previously made training look like an unbounded memory leak:

• build_dir loops forever on absolute paths (os.path.split("/") never yields ""); replaced with os.makedirs.
• float64 input crashes the float32 BatchNorm layers; cast to float32.

Feeds sciPENN the log_cp10k layers with its own normalize_total/log1p disabled, and maps its z-scored predictions back to log_cp10k using the training-protein mean/std, so outputs match the ground-truth space.

Also: arguments use default instead of example (they were None at runtime), add --seed, install requests for check_config, and stop listing run_and_check_output twice in test_resources.

Describe your changes

Checklist before requesting a review

• [ x] I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • [ x] New functionality
  • [x ] Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• [x ] CI Tests succeed and look good!

[benjaminfreyuu]

@VladimirShitov Please also review if time

[VladimirShitov]

In general looks good, but please take a look at the comments. You mentioned that ther ewas a problem with this method previously, what was it and how is it solved now?

[VladimirShitov]

If it is not model-specific, there is no need: HVGs are part of input data

[VladimirShitov]

Why such a weird format and not just 10000?

[VladimirShitov]

Never use one letter variable names please

[VladimirShitov]

Also please fix the CI failing

[benjaminfreyuu]

Thanks @VladimirShitov — addressed all the comments (viash test 2/2):

• CI failure: it was the torch install — torch==1.13.1 --index-url https://download.pytorch.org/whl/cu117 made pip resolve all packages (incl. the flit_core build backend) from the pytorch index, which doesn't host them. Switched to building on openproblems/base_pytorch_nvidia:1 with just pip install --no-deps sciPENN==0.9.6 (the base ships a working torch/scanpy/anndata stack + the openproblems core helper the test harness needs). sciPENN runs fine on it.
• select_hvg: removed — and for the same reason (min_cells/min_genes are data preprocessing, not model hyperparameters, and aren't exposed by the other methods) I dropped those two as well. Their values are hardcoded in the script so behaviour is unchanged.
• int(r'10000'): that was the auto-generated viash config inject block; replaced with a hand-written VIASH START using plain literals.
• One-letter variables: renamed all of them.

On the "previous problem": the July diagnosis ("training memory leak, needs GPU + highmem") was wrong. The real causes, now fixed:

• sciPENN 0.9.6's build_dir infinite-loops on absolute paths (os.path.split("/") never returns "", so it appends "/" forever until the process is OOM-killed) — that was the "memory leak". Fixed by monkeypatching build_dir with os.makedirs. Actual training is ~2 s / ~430 MB on the test data.
• float64 input crashed sciPENN's float32 BatchNorm → the script casts inputs to float32.
• The config used example: instead of default:, so args arrived as None at runtime.
• sciPENN emits per-batch z-scored proteins; the script now feeds log_cp10k and inverts predictions with the training-protein mean/std, which dropped RMSE from 3.48 to 1.27.