[runtime] PyTorch model MLIR JIT decorator

[adam-smnk]

Adds a jit decorator to compile and execute torch models using MLIR.

[adam-smnk]

Module becomes pretty ambiguous between torch and MLIR.
Now naming leans into model when referring to PyTorch and module is left for MLIR context.

[adam-smnk]

@rolfmorel You were right. Decorator hiding the original class makes it less useful.
I dug around torch.nn.Module and managed to hook into its existing callback to redirect model's execution through the custom JIT logic.

For completion, I should mention that there's an option to use torch.compile and provide it with a custom backend.
This requires that the returned callable has the same signature as the model's forward function which is not feasible if MLIR transformations modify the compilated function's signature.

I think the above could be possible with some argument normalization step between the model and MLIR. But not sure if it's worth going that far just yet.