[Torch] Fix dtype handling for modules with integer parameters

[masahi]

This fixes the interesting typing problem raised in #6300.

In Torchscript, parameters of operations like conv are always accessed via prim::GetAttr nodes. For example, when we see

 %input : Float(1, 6, 4) = aten::_convolution(%input.1, %6, %7, %9, %11, %13, %14, %16, %17, %18, %19, %20)

The input weight %6 is accessed in this way:

%3 : Module = prim::GetAttr[name="conv"](%self.1)
%6 : Tensor = prim::GetAttr[name="weight"](%3)

The problem is, Torch cannot figure out the correct type of GetAttr nodes. So when we visit aten::_convolution to get its input types, we have to assume that this is an untyped tensor and get annoying warnings Untyped Tensor found, assume it is float.

This hasn't been a big issue so far because parameters are usually float anyways. But #6300 brought a use case where there are integer parameters as well as float ones. Changing default_dtype to int doesn't solve it.

So I added a workaround when we try to get the dtype of GetAttr nodes. For every GetAttr node there is a corresponding parameter from the original PyTorch module with known, correct dtype. And for every PyTorch parameter tensor we have a corresponding Relay Var (via convert_params(...) function). So inside _get_input_types when we find GetAttr node, we return the dtype of corresponding Relay Var, instead of returning default_type.

This is the solution I came up with minimal change. It fixes the problem, but I feel it is a bit hacky. Please let me know if there are better ways to handle this. The test case I added is a minimal reproduction of the issue raised in #6300.

As a bonus, there would be no more annoying warnings Untyped Tensor found... when working with traced models.

please review @siju-samuel @t-vi

[siju-samuel]

LGTM
WARNING:root:Untyped Tensor found, assume it is float32 -> This annoying warning is gone. Thanks.

[siju-samuel]

Thanks @masahi. This PR is merged.