[Triton] MXFP8 GEMM and A8W4 MOE optimization for DSV4

[k50112113]

This PR is co-authored by @azaidy and @k50112113
This PR depends on ROCm/aiter#3292

This PR includes:

1. Added MXFP8 GEMM and related quant/fusions for DSV4, for example fuse flatten with mxfp8 quant after einsum.
2. Enabled A8W4 Triton MOE (fused with activation and mxfp8 downcast into MOE GEMM1)
3. Optimized Triton Topk and routing for the A8W4 MOE

We see +3~7% e2e performance bump across different conc, need to set

export ATOM_FP8_BLOCKSCALE_USE_MXFP8=1
export ATOM_USE_TRITON_MOE=1

lm_eval gsm8k full test set

local-completions ({'model': '/data/deepseek-ai/DeepSeek-V4-Pro', 'base_url': 'http://localhost:8000/v1/completions', 'num_concurrent': 65, 'max_retries': 3, 'tokenized_requests': False}), gen_kwargs: ({}), limit: 2000.0, num_fewshot: 3, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.9515|±  |0.0059|
|     |       |strict-match    |     3|exact_match|↑  |0.9522|±  |0.0059|

Total token throughput (toks/s):

Conc	Baseline	This PR, export ATOM_FP8_BLOCKSCALE_USE_MXFP8=1, and export ATOM_USE_TRITON_MOE=1	speedup
4	417.48	439.87	1.054
8	788.78	821.97	1.042
16	1427.29	1473.69	1.033
32	2507.72	2593.96	1.034
64	4088.5	4367.02	1.068