[TLX] Interleave TMA stores across MMA groups in Blackwell GEMM epilogue

[htyu]

Summary:
Add an interleaved epilogue mode to the Blackwell warp-specialized GEMM
kernel. When INTERLEAVE_EPILOGUE=1, the epilogue alternates TMA stores
between MMA group 0 and group 1 instead of draining each group
sequentially. This overlaps the TMA store latency of one group with the
TMEM read of the other, improving memory throughput on store-bound shapes.

The interleaved path is enabled by default for GPU-saturated shapes and
for tall-M shapes with small K (low arithmetic intensity). It is disabled
for Split-K configs (which use atomic reductions) and for the tall-M
high-arithmetic-intensity path with BLOCK_K=128.

Autotuning is also extended to explore INTERLEAVE_EPILOGUE in {0, 1},
with config pruning updated to filter invalid combinations (interleave
requires NUM_MMA_GROUPS == 2 and SPLIT_K == 1).

Perf on B200 (tflops):

                         aten      tlx_matmul_ws
           (M, N, K)  matmul   before    after   delta
  (8192, 8192, 8192) 1142.09  1168.46  1182.49  +1.2%
 (3159809, 384, 384)  647.23   664.21   644.05  -3.0%
(1152, 12800, 32768) 1124.60  1076.00  1069.98  -0.6%
  (1024, 256, 16384)  363.73   209.39   209.72  +0.2%
  (560849, 512, 896)  889.47   898.91   938.96  +4.5%
 (589824, 512, 2048)  915.12   959.46   959.39  -0.0%
 (1152, 65536, 1024) 1071.84   926.53   962.52  +3.9%
  (8192, 4608, 6144) 1170.41  1176.01  1195.01  +1.6%
(16384, 11264, 5632) 1089.06  1132.88  1149.37  +1.5%
  (8192, 8192, 2048) 1193.88  1141.82  1162.53  +1.8%
             average  960.74   935.37   947.40  +1.3%

Differential Revision: D94608909

[meta-codesync]

@htyu has exported this pull request. If you are a Meta employee, you can view the originating Diff in D94608909.