Adding shape information for Torch MLIR LTC Backend

[henrytwo]

After #725 lands, the next goal is to try to add shape information to the output MLIR.

lazy::Node contains tensor shape information (retrieved by calling shapes()); however, this information is not retained when it's converted to a jit::Node. (This shows that the only thing extracted from the lazy::Node during lowering is its symbol)

It also looks like there is nothing on the MLIR generation side for adding shape information (as far as I could see).

I don't think there is anything on the JIT side that enables storing tensor shape information, which might make this more difficult. My immediate thought is to try to map lazy::Node -> MlirOperation to insert shape information, but that doesn't feel like the most proper way to do this.

@silvasean, do you have any thoughts or ideas on how to proceed with this?

cc: @antoniojkim @ke1337

[silvasean]

c10::TensorType seems to support having a shape specified, so any jit::Value's of TensorType should be able to hold a shape. I think the path forward here is for upstream to create the jit::Value's with more precise TensorType's.

[henrytwo]

Ah that's good to hear -- I'll investigate what we can do upstream

[henrytwo]

I opened an issue on the PyTorch side here: pytorch/pytorch#75217. Through some testing it seems like on the torch-mlir side we already generate MLIR with the correct shape information, which is nice to see.

[henrytwo]

Closing this issue now that we have a PR open: #742