feat: native zero-copy exchange for all Zarr memory layouts

[kylebarron]

As explored before, such as in #29, we need to think carefully about our data layouts so they can be exchanged to Python with minimal overhead or copying.

This PR takes the approach of designing DecodedArray, a custom holder for ArrayBytes and which implements FromArrayBytes. This means DecodedArray can be used directly as a result type from operations like Array.retrieve_chunk.

The DecodedArray is actually an enum over four concrete types (essentially the ArrayBytes variants):

• Tensor: primitive type
• Variable: variable length strings and bytes
• Masked tensor: Tensor + a validity mask
• MaskedVariable: Variable + a validity mask

These each become Python types that the user can use isinstance to determine which they have.

Each of these types expose different options for zero copy conversion:

• Tensor: buffer protocol, numpy, dlpack, arrow
• Variable: arrow
• MaskedTensor: numpy masked array, arrow
• MaskedVariable: arrow

Closes #29