Update anndata (version and import source of `write_elem`)

[tcompa]

See https://anndata.readthedocs.io/en/stable/release-notes/index.html

[jluethi]

Would be great to evaluate this. I think the write_elem has now also moved out of the experimental part, thus maybe we can ignore anndata's torch interactions with that. I'll test this briefly for the md converter, where write_elem in experimental has been making some issues.

[jluethi]

Actually, the more important part will be this: Let's switch to using the write_elem function as ad._io.specs.write_elem, instead of importing it from experimental. It has worked in my tests for the MD, let's try also on the Fractal side.

The big benefit: ad._io.specs does not import torch => no more anndata worries about torch.

See: https://github.com/jluethi/fractal-faim-hcs/issues/1

[tcompa]

Let's switch to using the write_elem function as ad._io.specs.write_elem, instead of importing it from experimental. It has worked in my tests for the MD, let's try also on the Fractal side.

The big benefit: ad._io.specs does not import torch => no more anndata worries about torch.

Fully agreed.

For the record:

• The write_elem function is exactly the same, since the experimental module simply imports from _io.specs (see https://github.com/scverse/anndata/blob/main/anndata/experimental/__init__.py, where it also imports torch).
• The _io module is still "experimental" and could change without notice, so I'm setting anndata version to be anndata = ">=0.8.0,<=0.9.1. We can incrementally include other versions after testing.