[Attention Performance] Flash Attention performance get to 80%~90% of XeTLA

[Dewei-Wang-sh]

we aim to get 80%+ of XeTLA
use python/tutorials/06-fused-attention.py as the test case.

• [#0 Attention Performance] get xetla's attention perf data and tiling strategy  #912
• [#1 Attention Performance] raise non-block pointer to block pointer #913
• [#2 Attention Performance] match attention workload when convert-triton-to-tritongpu-warp #914
• [#3 Attention Performance] add attention support to tritonintelgpu-distribute-to-warps #915
• [#4 Attention Performance] add attention support to tritonintelgpu-match-target-size #916
• [#5 Attention Performance] add attention support to convert-tritongpu-to-llvm #917
• [#6 Attention Performance] extend attention support for Causal = True #1102
• [#7 Attention Performance] compare triton optimizations to XeTLA counterpart optimization #1103
• [#8 Attention Performance] reduce ALU instruction and register spill in the loop body #1192

(batch head n_ctx d_head causal) on max1100
for fwd_1x2x1024x32_true, xetla median is 4.7tflops
for fwd_1x2x1024x32_false, xetla median is 4.6tflops
for fwd_4x48x1024x64_true, xetla median is 110tflops
for fwd_4x48x1024x64_false, xetla median is 65tflops

[Dewei-Wang-sh]

For case fwd_4x48x1024x64_false (batch, num_head, n_ctx, dim_head, causal), it can get 60% on GPU Max 1550.
end2end run with some hack code, but result data mismatch

[Dewei-Wang-sh]

fixed data mismatch;
only collect the perf data that flushes cache.
For case fwd_4x48x1024x64_false (batch, num_head, n_ctx, dim_head, causal), it can get 66% on GPU Max 1550.

[Dewei-Wang-sh]

need #1102 to close this umbrella issue.