feat: output anndata

[ilan-gold]

Fixes #1

See #1 (comment) for the plan / approach.

Currently dependent on the bug highlighted in scverse/anndata-rs#44 which may not be a bug but instead something that should be changed in python.

Todos:

• Resolve the obsm dataframe mess
• Clean up some of the AI output
• Relatedly, decide on how extras should be structured
• Check speed

[BenjaminDEMAILLE]

Heads-up on the conflict state, from a pass over the whole PR queue.

Two of the three conflicts against current main are mechanical and I verified a resolution locally:

• src/solo/mod.rs: pub mod adata; and pub mod cell_reads; are both additions, so both stay.
• src/junction/sj_output.rs: main still inlines the surviving-junction filter and sort inside write_sj_lines, while this branch factored the same logic into sorted_junctions(params). The helper is equivalent (same filter, same sort key), so this branch's version wins and main's inline block goes.

The third is not mechanical and is yours to decide: src/solo/count.rs has two ~110-line hunks where this branch rewrote the matrix writer to take a cells slice, and main reworked the same function's record folding (1MM_multi resolution off exact_count_snapshot, std::mem::take of the recorder). Picking one shape over the other changes what the anndata writer receives, so I left it alone rather than guessing.

No commits were pushed to this branch.