Support reading dense tensors from sparse arrays

[gsakkis]

As mentioned in the #167 description, the TensorKind abstraction allows decoupling the TileDB array flavor (dense or sparse) from the generated tensor kind. This PR enables generating dense tensors from sparse TileDB arrays by passing tensor_kind=TensorKind.DENSE to the ArrayParams of a sparse array.

The actual implementation is contained in a single commit (56599f6) that:

• generalizes the TensorSchemaFactories mapping to depend on the array flavor (in addition to the tensor kind).
• adds SparseToDenseTensorSchema factory that maps sparse arrays generated by SparseTensorSchema to (dense) Numpy arrays.
• makes MappedTensorSchema pickleable (required for num_workers>0)

[shortcut-integration]

This pull request has been linked to Shortcut Story #21303: Support reading dense tensors from sparse arrays.