Update scRNA tutorial

[flying-sheep]

Fixes #153

basic clustering tutorial notebook

• Seems like the manual annotation is already broken, so I removed it.
  If you have an idea how we can keep it stable, we could do that instead, but I couldn’t figure out which cluster assignments actually work.
• Since {eval}thing does not work with nb_execution_mode == 'off' executablebooks/MyST-NB#596 hasn’t been merged for ages, I’m installing it from git for now. It’s a good way to reference not-too-stable numbers from text without having to manually keep an eye on things

Other notebooks

• fixes scipy compat by changing csr_matrix.{A→toarray()}
• fixes the backed mode tutorial’s dotplot, which used dendrogram=True: sc.tl.dendrogram no longer(?) works in backed mode scanpy#3199

Conda setup

Some of the conda setup does no longer work, namely use-only-tar-bz2, so I needed to disable it. Let’s fully fix the setup in another PR, this PR makes it work again which was hard enough.

We should probably go away from conda in the first place. Installing pip packages into a conda environment is an undefined operation, all version bounds for the conda packages become meaningless when doing that. uv could be enough, otherwise we could go pixi or so.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[grst]

Seems like the manual annotation is already broken, so I removed it.
If you have an idea how we can keep it stable, we could do that instead, but I couldn’t figure out which cluster assignments actually work.

In scverse/scanpy#2014, we figured out that setting

export NUMBA_CPU_NAME=generic

makes the clustering results more stable (at least in the conditions I tested there) by only relying on basic CPU features. Would still need to redo it once but then hopefully it wouldn't change anymore in the future.

[grst]

LGTM! Thank you for suffering for us!

I would be in favor of giving it a shot to try out the reproducible clustering with the numba variable as manual clustering is still a commonly used approach.

[grst]

And not sure why we went with conda in the first place, maybe because of compatibility with mybinder?

[flying-sheep]

let’s restore the manual clustering in another PR, I don’t feel like going through all clusters and assigning labels manually.