Detect sub-group shuffles with sliced dimensions

[victor-eds]

#2531 fails to detect layout conversions with sliced dimensions as shuffles. This way, code like:

#blocked = #triton_gpu.blocked<{sizePerThread = [1, 16], threadsPerWarp = [16, 1], warpsPerCTA = [1, 2], order = [0, 1]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [16, 1], threadsPerWarp = [1, 16], warpsPerCTA = [1, 2], order = [0, 1]}>
#sliced = #triton_gpu.slice<{dim = 1, parent = #blocked}>
#sliced1 = #triton_gpu.slice<{dim = 1, parent = #blocked1}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-warp" = 16 : i32} {
  tt.func @test_f32(%arg0: tensor<32xf32, #sliced>) -> tensor<32xf32, #sliced1> {
    %0 = triton_gpu.convert_layout %arg0 : tensor<32xf32, #sliced> -> tensor<32xf32, #sliced1>
    tt.return %0 : tensor<32xf32, #sliced1>
  }
}

won't be lowered to shuffles, but to SLM memory accesses needing synchronization.

[victor-eds]

Not really what this ticket was issued for, but a nice improvement we will build upon when implementing this:
#2573

[victor-eds]

Pushed first PR extending this pass' capabilities. Will need to rebase to adapt to new LinearLayout interface. Working on last PR to enable the feature.

[victor-eds]

After #2573, we optimize all layout conversions in the FA benchmark runs. As making this more generic would take too much time, I'd rather spend time somewhere else and call this done as its original purpose (optimizing the cases we run into) is done.