Generalize FoldMaskedTransferRaw and add FoldTransferReadOfEmptyTensor (#24301)

* Generalizes FoldMaskedTransferRAW from (masked, masked) to (unmasked,
unmasked), (unmasked, masked), (masked, unmasked).
* Adds pattern to fold transfer_read(tensor.empty)) -> ub.poison
* This allows intermediary index tensors to be folded after
vectorization.
* The test pipeline_vector_distribute_reduction_gfx942.mlir needed to be
corrected. The empty tensors are now folded, but the test was wrong,
online_attention should not have had empty tensors as operands to begin
with. All passes that create online_attention fill the operands with
either 0 or -1. So we do that here as well.

Fixes #24294

----

Assisted-By: Claude Opus 4.6