Infer matmul dimension roles to compute vectorization

[jacobhinkle]

This PR does the following:

1. Rename RolesMap to TensorRolesMap and introduce DimRolesMap which is a mapping from ValGroup to MatmulDomain.
2. Compute a canonical dim ordering on ValGroups based on allocation domains of inputs and outputs. This is used to compute vectorization properly but can be used for canonicalization of loop domains in scheduleMatmul in a future PR.
3. Properly infer vectorization for every operand, epilogue input, and output based on canonical dim ordering.

This is in preparation for further generalization to accomodate multiple MmaOps in a single Fusion.

Fixes #2169.

[jacobhinkle]

!build

[jacobhinkle]

!build

[jacobhinkle]

!build

[jacobhinkle]

!build --diff

[jacobhinkle]

I think the codediffs are just due to the uncommented tests.

[zasdfgbnm]

What if my mma output is (M=1024, N=1024), and my epilogue is:

T1[1024, 128, 2, 2', 2"] = view(mma_output);
T2[1024, 128, 2', 2, 2"] = transpose(T1);
T3[1024, 1024] = view(T2);
output T3

[jacobhinkle]

What if my mma output is (M=1024, N=1024), and my epilogue is:

T1[1024, 128, 2, 2', 2"] = view(mma_output);
T2[1024, 128, 2', 2, 2"] = transpose(T1);
T3[1024, 1024] = view(T2);
output T3

In that case, I would expect T3 to not be mapped as an OUTPUT_D tensor, in which case we currently would refuse to segment this because we do not allow tensors without known roles. In the future, we could try and allow things like this by loosening the restriction that output dims must exact map to M, N, K or Batch dims. Then we could do something like what @protonu did for #2315; Use unmapped dims to traverse backward through producers to find mapped dimensions and determine vectorization. In that case we could accept since the inner dimension of T3 corresponds to a shuffled N dimension, and we'd select vectorization width of 2 for T3. This particular example would also require us to accept epilogues with ViewOp in them, which we currently block, but it's at least feasible I think.

[zasdfgbnm]

Oh, I see. We are effectively rejecting view ops inside the fusion. And this rejection saves us from having to use the vectorization helper.

[jacobhinkle]

!build