Point scprint's triton at its own ptxas

[rcannood]

Describe your changes

scprint died on every one of its 12 tasks in the last full run with

RuntimeError: Triton only support CUDA 10.0 or higher

which is triton's message for "I do not recognise this CUDA version". The image pins torch 2.2.0+cu121 and triton 2.2.0, whose ptx_get_version() only maps CUDA 10, 11 and 12, while openproblems/base_pytorch_nvidia:1 now ships CUDA 13.1 at /usr/local/cuda. Triton finds that ptxas first and gives up the moment simpler_flash JITs a kernel.

Triton 2.2 bundles its own CUDA 12.3 ptxas, so pointing TRITON_PTXAS_PATH at it is enough.

Verified on a de.NBI T4 by re-running the failing task's own work dir with that variable set: it completes and writes a valid embedding for all 39176 cells.

Also drops the stray commented # docker_run_args: --gpus all under runners:, which had odd indentation and did nothing.

Bumping scprint/torch/triton to something that understands CUDA 13 is the real answer, but that pulls on simpler_flash and the lamin setup, and this is enough to get the method scoring again.

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!