Add additional matching logic to MMGroupQuantRewriterPass

[Max191]

This adds additional matching logic on tensor.view ops required for using MMGroupQuantRewriterPass in the llama.turbine model (used here stellaraccident/llama.turbine#2).

[stellaraccident]

Aside from the test failures, I think I need to see the IR we are trying to match. But this seems like it might be tricky to keep working.

[stellaraccident]

Remove

[Max191]

Aside from the test failures, I think I need to see the IR we are trying to match. But this seems like it might be tricky to keep working.

The IR this is matching looks like this:

    %_params.blk.0.ffn_gate.weight = util.global.load @_params.blk.0.ffn_gate.weight : tensor<3200x8640xf32>
    %152 = torch_c.from_builtin_tensor %_params.blk.0.ffn_gate.weight : tensor<3200x8640xf32> -> !torch.vtensor<[3200,8640],f32>
    %int8640 = torch.constant.int 8640
    %int3200_171 = torch.constant.int 3200
    %153 = torch.prim.ListConstruct %int8640, %int3200_171 : (!torch.int, !torch.int) -> !torch.list<int>
    %154 = torch.aten.view %152, %153 : !torch.vtensor<[3200,8640],f32>, !torch.list<int> -> !torch.vtensor<[8640,3200],f32>
    %int1_172 = torch.constant.int 1
    %int0_173 = torch.constant.int 0
    %155 = torch.prim.ListConstruct %int1_172, %int0_173 : (!torch.int, !torch.int) -> !torch.list<int>
    %156 = torch.aten.permute %154, %155 : !torch.vtensor<[8640,3200],f32>, !torch.list<int> -> !torch.vtensor<[3200,8640],f32>
    %int3200_174 = torch.constant.int 3200
    %157 = torch.prim.ListConstruct %4, %int3200_174 : (!torch.int, !torch.int) -> !torch.list<int>
    %158 = torch.aten.view %151, %157 : !torch.vtensor<[1,?,3200],f32>, !torch.list<int> -> !torch.vtensor<[?,3200],f32>
    %159 = torch.aten.mm %158, %156 : !torch.vtensor<[?,3200],f32>, !torch.vtensor<[3200,8640],f32> -> !torch.vtensor<[?,8640],f32>

This change basically just rewrites the parameter to the viewed type and elides the view (%154) as long as the view is just swapping the shape sizes, and the group size evenly divides the new inner shape.

This pass was already pretty specific to llama, especially since it was matching for a specific parameter name, so I figured that this should be okay for now. Maybe the view rewrite could be moved to its own pass, but this just seemed like the easiest path forward for me, and I like the idea of having fewer passes to call after exporting.