Revert tensor.cast to flow reshape conversion

[IanWood1]

Revert tensor.cast to flow.tensor.reshape conversion because this is occurring too early. This happens as early as GlobalOptimization since Flow CanonicalizerPass is used there. Also, stop cloning tensor::CastOps into dispatches.

Reverted pr: #18351 Original Issue: #18229

[IanWood1]

Okay, it appears like everything is working. I just need to fix up dispatch_linalg_on_tensors.mlir

[hanhanW]

Is the plan adding shape inference to tensor.expand/collapse_shape's canonicalization patterns?

cc @nirvedhmeshram who landed the patch

[IanWood1]

Yes, I have a WIP pr llvm/llvm-project#112265. Also, It looks like collapse_shape already has this pattern

[nirvedhmeshram]

LGTM, yes converting to reshape early was working around an issue of tenor.casts in dispatches at the expense of shape inference, so this way we can do that shape inference on tensors and by not cloning the casts in the dispatches the original issue wont happen. Thanks!

[IanWood1]

Given CI runs against onnx ops (https://github.com/iree-org/iree/actions/runs/11336536464/job/31526991018?pr=18772) I think this should be good to merge?

[nirvedhmeshram]

Given CI runs against onnx ops (https://github.com/iree-org/iree/actions/runs/11336536464/job/31526991018?pr=18772) I think this should be good to merge?

The CI doesn't really test the full coverage of onnx models but if the IR in the original issue is still compiling, I think this is okay to merge.

[IanWood1]

iree-compile hit an assertion due to the elided dense attrs (after compilation) which I assume means it works

[nirvedhmeshram]

iree-compile hit an assertion due to the elided dense attrs (after compilation) which I assume means it works

Yes, that old IR had that issue so you are good to go.