[PyTorch] Add aten::embedding_bag

[shingjan]

This PR intends to add aten::embedding_bag for the pytorch frontend. Note that the implementation of aten::numel is also changed under the condition that any input that can be evaluated to a constant value will be evaluated at compile time.

Co-authored-by: Masahiro Masuda masahi@129@gmail.com

cc: @masahi @vinx13

[shingjan]

@masahi This implementation still failed the test case with 1D input as the offsets are parsed into:

free_var %input2: Tensor[(2), int64];
%input2

which cannot be folded to constant at compile time even this 1D case is actually identical to 2D case:

    # 1D case
    input = torch.tensor([2, 2, 2, 2, 4, 3, 2, 9])
    offsets = torch.tensor([0, 4])
    # which is equivalent to  
    input = torch.tensor([[2, 2, 2, 2], [4, 3, 2, 9]])

My take is that we can support 2D case here and see if there could be any workaround for the 1D cases like above. Note that sparse and per_sample_weights are also taken care of.

[shingjan]

closed with #12993