[XPU][TritonGPUToLLVM] Avoid bank conflicts in sub-group transposes

- Store the whole matrix using SIMD block stores for each row leaving
  a single garbage item at the end of the row so each row has
  `sub_group_size + 1` elements
- Load each row with vector loads

By introducing this garbage item at the end of each row, we ensure matrix
loading avoid bank conflicts as the offset between the position loaded by
work-item `i` and `i+j` is `N * (sub_group_size + 1)` (assuming `sub_group_size`
banks).

Signed-off-by: victor-eds <[email protected]>

test/Conversion/intel/intel-allocate-shared-memory.mlir

@@ -24,7 +24,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
 // Check scracth memory configuration for different sub-group transpose-like layout conversions.
 
 // CHECK-LABEL: module attributes
-// CHECK-SAME: triton_gpu.shared = 512 : i32
+// CHECK-SAME: triton_gpu.shared = 544 : i32
 module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-warp" = 16 : i32} {
   tt.func @test_f16(%arg0: tensor<16x16xf16, #blocked>) -> tensor<16x16xf16, #blocked1> {
     %0 = triton_gpu.convert_layout %arg0 : tensor<16x16xf16, #blocked> -> tensor<16x16xf16, #blocked1>
@@ -40,7 +40,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
 // Check scracth memory configuration for different sub-group transpose-like layout conversions.
 
 // CHECK-LABEL: module attributes
-// CHECK-SAME: triton_gpu.shared = 1024 : i32
+// CHECK-SAME: triton_gpu.shared = 1088 : i32
 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<16x16xf32, #blocked>) -> tensor<16x16xf32, #blocked1> {
     %0 = triton_gpu.convert_layout %arg0 : tensor<16x16xf32, #blocked> -> tensor<16x16xf32, #blocked1>
@@ -56,7 +56,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
 // Check scracth memory configuration for different sub-group transpose-like layout conversions.
 
 // CHECK-LABEL: module attributes
-// CHECK-SAME: triton_gpu.shared = 32768 : i32
+// CHECK-SAME: triton_gpu.shared = 34816 : i32
 module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 8 : i32, "triton_gpu.threads-per-warp" = 16 : i32} {
   tt.func @test_f32(%arg0: tensor<128x64xf32, #blocked>) -> tensor<128x64xf32, #blocked1> {
     %0 = triton_gpu.convert_layout %arg0 : tensor<128x64xf32, #blocked> -> tensor<128x64xf32, #blocked1>