Improve ergonomics of `FillValue` API by storing `DataType`

Right now our `FillValue` class is a very minimal wrapper of upstream Zarrs [`FillValue`](https://docs.rs/zarrs/latest/zarrs/array/struct.FillValue.html), which holds only a raw `Vec`.

Perhaps our `FillValue` should also hold a `DataType` so that it can be smarter about the user-facing API it presents?

Like it could have a `to_numpy` API that would then create a `np.uint8(value)` or `np.float32(value)` etc.

Then the user wouldn't have to manually convert raw bytes into a specific data type's fill value.

This would change our Python-facing `FillValue` to be _fully generic_ to being _tied to a specific array's data type_, but I think that's an improvement in ergonomics.

The entry point is the Python-facing FillValue wrapper, with upstream zarrs FillValue as the reference; first review how its raw Vec and DataType are currently exposed. Done means a FillValue can be associated with a DataType and provide the proposed typed to_numpy behavior without requiring manual raw-byte conversion.

Written by the indexing model from the issue text.