Bump cellmapper_scvi's base image and align both obsm keys

[rcannood]

Describe your changes

cellmapper_scvi fails on every run on current main:

KeyError: 'mod2_pred'

This is my fault -- #23 was wrong. I checked cellmapper's GitHub main instead of the version the image actually installs. Sorry for the churn.

What's really going on

cellmapper changed how map_obsm names its output key, and the two cellmapper components were on different sides of that change because of their base images:

release	builds
0.2.2 / 0.2.3	f"{key}_{prediction_postfix}"
0.2.4+	f"{key}{prediction_postfix}"

cellmapper_scvi pinned openproblems/base_pytorch_nvidia:1.0.0, which is on Python 3.10 -- and 0.2.4+ requires ≥3.11 (0.2.6 requires ≥3.12), so pip there could only ever resolve 0.2.3:

ERROR: Ignored the following versions that require a different python version:
  0.2.4 Requires-Python >=3.11; 0.2.5 Requires-Python >=3.11; 0.2.6 Requires-Python >=3.12

On 0.2.3 the "_pred" postfix yields mod2__pred, hence the KeyError.

The change

Rather than pinning the component down to 0.2.3, bump the base image -- :1.0.0 was the only non-sliding base image left in the repo, everything else already uses :1, and :1 is on Python 3.12:

• cellmapper_scvi: base image :1.0.0 -> :1. The existing prediction_postfix="_pred" is then correct, so the script is untouched.
• cellmapper_linear: switch to the same "_pred" postfix, so both components read obsm["mod2_pred"].
• Both: require cellmapper>=0.2.6. An open >=0.2.2 makes the correct key depend on when the image was built, which is how this slipped through twice.

The bumped image also picks up current versions of everything else:

cellmapper-0.2.6  scvi-tools-1.5.0.post1  muon-0.1.9
lightning-2.6.5   mudata-0.3.10           torchmetrics-1.9.0

Verified on a de.NBI Tesla T4

Rebuilt both images from scratch (---setup cb) and ran them, sampling nvidia-smi --query-compute-apps once a second:

cellmapper_scvi     EXIT=0  42s  peak 12% util  165 MiB  CUDA proc 17/39  -> USES GPU
  prediction shape=(163, 134) method_id='cellmapper_scvi' CONTRACT=OK

cellmapper_linear   EXIT=0  23s                          (no gpu label)   -> CPU, as expected
  prediction shape=(163, 134) method_id='cellmapper_linear' CONTRACT=OK

Both images confirmed on cellmapper 0.2.6. cellmapper_scvi now produces a valid prediction and genuinely uses the GPU -- as far as I can tell it has never produced a score before this.

Found while checking on de.NBI which GPU-labelled methods actually use the GPU.

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!