API Compat fix on max_pool2d_with_indices_()

[brucekimrokcmu]

This PR addresses the following error to resolve a number of test cases:

1 : MaxPool2dWithIndicesFullSizeKernelModule_basic
max_pool2d_with_indices(): incompatible function arguments. The following argument types are supported:
    1. (self: pi.mlir._mlir_libs._pi_mlir.Tensor, kernel_size: pi.mlir._mlir_libs._pi_mlir.AnyTorchListOfTorchIntValue, stride: pi.mlir._mlir_libs._pi_mlir.AnyTorchListOfTorchIntValue = [], padding: pi.mlir._mlir_libs._pi_mlir.AnyTorchListOfTorchIntValue = [0, 0], dilation: pi.mlir._mlir_libs._pi_mlir.AnyTorchListOfTorchIntValue = [1, 1], ceil_mode: pi.mlir._mlir_libs._pi_mlir.Torch_BoolValue = False, *, loc: mlir.ir.Location = None, ip: mlir.ir.InsertionPoint = None) -> Tuple[pi.mlir._mlir_libs._pi_mlir.Tensor, pi.mlir._mlir_libs._pi_mlir.Tensor]

Invoked with: Tensor(<block argument> of type '!torch.tensor' at index: 0); kwargs: kernel_size=[Torch_IntValue(%0 = "torch.constant.int"() {value = 4 : i64} : () -> !torch.int), Torch_IntValue(%1 = "torch.constant.int"() {value = 4 : i64} : () -> !torch.int)], stride=1, padding=0, dilation=1 

Essentially, a recursive template function generates 16 different combinations ofPyAnyTorchListOfTorchIntValue and PyTorch_IntValue for four arguments (kernel_size, stride, padding, dilation) to be accepted for max_pool2d_with_indices_ Ops.

Note that to avoid segfault, max_pool2d_with_indices() redefines default loc and ip before all casted arguments are passed into max_pool2d_with_indices_() ops function.

[123epsilon]

The purpose of this PR is to address several test cases for max_pool2d_with_indices that require accepting different combinations of a list versus an int for parameters such as kernel_size. Since there are a lot of combinations (16 for 4 parameters) we generate the appropriate bindings via recursive template instantiation. There is one downside in the current implementation: we do not have default arguments for the arguments to the Op.

This passes an additional 6 tests

[makslevental]

the red x here is because of whate we discussed offline? missing torch-mlir wheel?

edit: indeed it is

[brucekimrokcmu]

Allowed max_pool2d and max_pool2d_with_indices to accept kwargs with default values to let users omit some of the kwargs in use. This passed one more test case, but essentially allowed freedom in passing kwargs or not.

[123epsilon]

@brucekimrokcmu it seems like you rebased against main but your PR is trying to change the torch-mlir submodule, can you verify whether your local branch is at the most recent torch-mlir commit in externals/torch-mlir it should be 1e468e8, the commit history looks a little strange to me