add var_len case for benchmark_mla_decode

[XiaobingSuper]

This PR is about adding var_len case for MLA benchmark.

[XiaobingSuper]

@tridao , for var_len case, I find FA3 has a big performance gap compared with FlashMLA(tested on H100):

total_seqlens=130864, mean_seqlens=1022, max_seqlen=2358
Seqlen = 2560, FA3 time: 131.5 us, 1418 GB/s, 277 TFLOPS/s
Seqlen = 2560, FlashMLA time: 107.6 us, 1732 GB/s, 339 TFLOPS/s
Arithmetic intensity: 195.5
Ideal time: 56 us
total_seqlens=274227, mean_seqlens=2142, max_seqlen=4696
Seqlen = 4864, FA3 time: 211.9 us, 1659 GB/s, 361 TFLOPS/s
Seqlen = 4864, FlashMLA time: 175.2 us, 2007 GB/s, 436 TFLOPS/s
Arithmetic intensity: 217.3
Ideal time: 105 us
total_seqlens=539326, mean_seqlens=4213, max_seqlen=9155
Seqlen = 9216, FA3 time: 363.3 us, 1808 GB/s, 413 TFLOPS/s
Seqlen = 9216, FlashMLA time: 292.9 us, 2243 GB/s, 513 TFLOPS/s
Arithmetic intensity: 228.7
Ideal time: 196 us
total_seqlens=1040384, mean_seqlens=8128, max_seqlen=17609
Seqlen = 17664, FA3 time: 664.7 us, 1857 GB/s, 436 TFLOPS/s
Seqlen = 17664, FlashMLA time: 505.0 us, 2444 GB/s, 574 TFLOPS/s
Arithmetic intensity: 234.8
Ideal time: 368 us
total_seqlens=2123978, mean_seqlens=16593, max_seqlen=37079
Seqlen = 37120, FA3 time: 1398.0 us, 1776 GB/s, 423 TFLOPS/s
Seqlen = 37120, FlashMLA time: 964.3 us, 2574 GB/s, 613 TFLOPS/s
Arithmetic intensity: 238.3
Ideal time: 741 us
total_seqlens=4190053, mean_seqlens=32734, max_seqlen=92105
Seqlen = 92160, FA3 time: 2515.6 us, 1933 GB/s, 464 TFLOPS/s
Seqlen = 92160, FlashMLA time: 1835.9 us, 2649 GB/s, 636 TFLOPS/s
Arithmetic intensity: 240.0
Ideal time: 1452 us
total_seqlens=8162046, mean_seqlens=63765, max_seqlen=147507
Seqlen = 147712, FA3 time: 5127.2 us, 1841 GB/s, 443 TFLOPS/s
Seqlen = 147712, FlashMLA time: 3511.0 us, 2688 GB/s, 648 TFLOPS/s
Arithmetic intensity: 240.9
Ideal time: 2817 us