(feat): remove dtype + fill val handling per chunk

[ilan-gold]

Calling to_native_dtype + __str__ came up as one of the only python-CPU-bound things when doing some benchmarking. My use-case is quite contrived (generating thousands of WithSubset objects) but I think it's probably worth investigating getting rid of these calls. Some observations:

1. I wonder if all getting the dtype and fill_val be wrapped up in just relying on https://docs.rs/zarrs/latest/zarrs/array/struct.Array.html#method.open and then using the values directly (there are probably other benefits of doing this) but I think this is a separate PR
2. Regardless, most of this refactor is around removing Basic anyway so that chunk handling is independent of the ability. I noticed that ChunkRepresentation requires ownership over its arguments which means we copy per-chunk. Not sure what would go into making that a reference, but it's no worse than the previous situation where I think we were generating copies repeatedly, but from PyO3 calling python

The benefit wasn't crazy ~5% but I think going in this direction is good (see point 1)

TODO:

• Understand our vlen test error messages / warnings re: what we support.