[Kernel] Fused MoE Config for Mixtral 8x22

[ywang96]

This PR adds fused MoE configs for Mixtral 8x22 on A100-80G & H100 with TP4 & 8 - Roughly 10% latency speedup for high batch sizes.

Attaching mean and median results from running benchmark_latency.py on A100-80G as a sanity check

TP4
bs 8
Avg latency: 5.787593524421876 seconds
50% percentile latency: 5.785140061285347 seconds

bs 16
Avg latency: 6.900521802188208 seconds
50% percentile latency: 6.910593919456005 seconds

bs 32
Avg latency: 7.437153746870657 seconds
50% percentile latency: 7.425246893661097 seconds

TP4 with config
bs 8
Avg latency: 5.4268179705676935 seconds
50% percentile latency: 5.4153253806289285 seconds

bs 16
Avg latency: 6.107941208345195 seconds
50% percentile latency: 6.10570535203442 seconds

bs 32
Avg latency: 6.893939550562451 seconds
50% percentile latency: 6.877172949025407 seconds

TP8
bs 8
Avg latency: 3.83942656442523 seconds
50% percentile latency: 3.8410149838309735 seconds

bs 16
Avg latency: 4.428924084718649 seconds
50% percentile latency: 4.421724954620004 seconds

bs 32
Avg latency: 4.767531445249915 seconds
50% percentile latency: 4.752460908377543 seconds


TP8 with config
bs 8
Avg latency: 3.7083894722008455 seconds
50% percentile latency: 3.7043474409729242 seconds

bs 16
Avg latency: 4.047313968433688 seconds
50% percentile latency: 4.0400408857967705 seconds

bs 32
Avg latency: 4.390694409639885 seconds
50% percentile latency: 4.380223051644862 seconds

[WoosukKwon]

@ywang96 LGTM. Thanks for the PR!