Add strand support to AnnDataset and fix deterministic_shift

[casblaauw]

Adds support for stranded datasets (so anndatas with indices (chr:start-end:strand), like gene data) while preserving support for non-stranded datasets (like ATAC data). I checked whether loading data works with stranded and unstranded data in-memory and not in-memory and with/without always_reverse_complement.

To do still:

• Check whether the downstream functions (crested object functions, like preds and explanations) still work. I looked at the code and nothing should need changes, but I haven't actually tested them yet - I can do that or someone else can quickly rerun an analysis notebook with this branch to see if everything works.

Also has two small bugfixes:

• get_embeddings() was trying to save per-gene embeddings in anndata.obsm (which is for per-pseudobulk observations). Changed this to .varm and tested that it works.
• the Crested.enhancer_design_* functions were deriving start and end from the var columns, but that's not robust: the the anndataset object always reads the full region string to get start and end, and indeed in my code (and I believe also if you use the expand region width function?) the regions and the start/end might be out of sync. Anyway, this was to get the model input size, so we can just use self.model.input_shape and not have to worry about it.

Edit: and a bigger bugfix: it fixes 'deterministic_shift', since it was broken before and didn't actually shift the sequences.

[casblaauw]

Now that we've swapped to always using augmented sequences, it makes handling the stranded logic easier and more robust, meaning all downstream functioning should work fine.
Since it passes all checks (which do also try the downstream options), I think this is ready for review and merging.

[LukasMahieu]

Tested, looks good. Can be merged, thanks for the fixes!