Refactor ta.fill_from_file

[pthomadakis]

Current ta.fill_from_file op expects an already existing sparse tensor as input and mutates its in place, generating memory side-effects.
To better adapt to mlir's semantics it would be more appropriate to refactor this op so that it is the own that generates the sparse tensor.
Example,
%0 = ta.sp_tensor_from_file(<filename>) : () -> !ta.sparse_tensor<...>
Similar to how tensor dialect's tensor.from_elements:
%0 = tensor.from_elements %a, %b, %c, %d, %e, %f : tensor<2x3xindex>