[Integrate] Drop llvm/llvm-project@b4c31dc revert.

[hanhanW]

It carries a cherry-pick fix that gets the operands from the adaptor:

• iree-org/llvm-project@8b88014

Changes:

• Update most lit tests to check vector.from_elements.
• Add unrolling patterns to the final conversion.
• Implement n-D vector::ToElementsOp lowering, which will be dropped after [mlir][vector] Add support for lowering n-D vector.to_elements op. llvm/llvm-project#156992 is landed. It should be added to all the backends, but somehow only AMDGPU backend needs the pattern. The other backends may address the issue via specialized tiling config + dropping vector unit dim patterns.

[hanhanW]

To repro: iree-compile --iree-hal-target-device=hip --iree-hip-target=gfx942 ~/repro.mlir

func.func @attention_2_1024_128_128_64_dtype_f16_f16_f16_f16(%query: tensor<2x1024x128xf16>, %key: tensor<2x128x128xf16>, %value: tensor<2x128x64xf16>, %scale: f32) -> tensor<2x1024x64xf16> {
  %result0 = tensor.empty(): tensor<2x1024x64xf16>
  %scale_f16 = arith.truncf %scale : f32 to f16
  %result1 = iree_linalg_ext.attention {
      indexing_maps = [affine_map<(batch, m, n, k1, k2) -> (batch, m, k1)>,
                       affine_map<(batch, m, n, k1, k2) -> (batch, k2, k1)>,
                       affine_map<(batch, m, n, k1, k2) -> (batch, k2, n)>,
                       affine_map<(batch, m, n, k1, k2) -> ()>,
                       affine_map<(batch, m, n, k1, k2) -> (batch, m, n)>]
}      ins(%query, %key, %value, %scale_f16: tensor<2x1024x128xf16>, tensor<2x128x128xf16>, tensor<2x128x64xf16>, f16)
      outs(%result0: tensor<2x1024x64xf16>) {
   ^bb0(%score: f32):
   iree_linalg_ext.yield %score : f32
 } -> tensor<2x1024x64xf16>
 return %result1: tensor<2x1024x64xf16>
}

[hanhanW]

@kuhar @MaheshRavishankar the CI is green. The cherry-picked commit is landed in upstream. I think we can land the PR, if you can help review the changes.

[Groverkss]

I don't know if flattening was the right call, this shouldve been an unrolling pattern. Everything else does unrolling.

[kuhar]

I think it's made explicit this is mostly a stopgap patch

[Groverkss]

fair enough