[ROCm] Enable Fused MLA Triton kernel for DeepSeekV3

[lcskrishna]

This PR introduces the concept of Fused MLA decode Triton kernel on ROCm. To use this feature one has to use the env variable : SGLANG_ROCM_FUSED_DECODE_MLA=1.

Triton Kernel authors: @juuso-oskari (Korhonen, Juuso), @Chi-Chu319 (Tianxing Wu) and @vgokhale (Gokhale Vinayak)

[lcskrishna]

cc: @sunway513

[lcskrishna]

@HaiShaw This is ready for review.

[HaiShaw]

@lcskrishna can you share some perf uplift or comparison here? Thanks!

[HaiShaw]

@saienduri can you help for a look, why all the CIs are skipped and unable to rerun?

[lcskrishna]

@lcskrishna can you share some perf uplift or comparison here? Thanks!

@HaiShaw Here are some basic benchmarks with DeepSeekV3 with this feature.

Without Fusion:

% python -m sglang.bench_one_batch --batch-size 4 --input 2048 --output 256 --model deepseek-ai/DeepSeek-V3 --tp 8 --trust-remote-code

Decode.  median latency: 0.03850 s, median throughput:    103.89 token/s
Total. latency: 10.656 s, throughput:    864.89 token/s

With this feature:

% SGLANG_ROCM_FUSED_DECODE_MLA=1  python -m sglang.bench_one_batch --batch-size 4 --input 2048 --output 256 --model deepseek-ai/DeepSeek-V3 --tp 8 --trust-remote-code

Decode.  median latency: 0.03581 s, median throughput:    111.71 token/s
Total. latency:  9.970 s, throughput:    924.39 token/s