minimize polymorphism in the codec pipeline?

[d-v-b]

we have a very expensive isinstance check in the codec pipeline:

zarr-python/src/zarr/codecs/bytes.py

Line 82 in 2a032a8

as_array_like = chunk_bytes.as_array_like()

. We need to remove this check, and I would suggest going further -- we should check if our strategy of supporting CPU and CUDA memory with the same codecs is compatible with getting the best performance for either type. I don't think supporting CUDA (used by a minority of users) is worth any measurable performance penalty for folks using CPU memory.