Handle large matrices and change default Seurat format

[jashapiro]

When testing AlexsLemonade/OpenScPCA-analysis#953, I found that the sum counts function doesn't work with very large matrices, namely unfiltered matrices. We probably don't really need to support that, but it wasn't too hard to do, so I added a bit of splitting and merging to avoid ever creating a single very large non-sparse matrix.

(There was a fun edge case I ran into when the matrix was one larger than a multiple of the chunk size...)

After some more Seurat testing, I also discovered that v5 is really the default for new objects, so we might as well make that our default too. So there are bunch of test changes related to that change in defaults.

[sjspielman]

This all looks good to me, and I like the strategy you used to split up matrices without actually splitting them up 🤙

The only comment I have is whether it's worth including a large matrix to test with, since our tests clearly didn't cover that exactly. Our test data here is 100 cells, and I guess we could "multiply" it by 50 (51?) and unique-ify barcodes along the way. That said, we don't actually test values of summed counts in the first place here, so this would require several more tests as a whole. I'm kind of on the fence for whether this is necessary, so I'm going to approve this and if you think it's a good idea to circle back to, then we can open an issue for future us.

[jashapiro]

I added an argument to set the cell chunk size, which made it easy to add some extra tests without making huge matrices.

[sjspielman]

🐰
i was typing for a tada emoji but accidentally typed ra... , so enjoy this rabbit who likes the new tests!