Fix CUTLASS DSL 4.6.0 compatibility

[wzhao18]

vllm bumped cute dsl version to 4.6.0, this PR fixes various compatibility issues introduced by the version bump.

Validated that all unit tests are now passing:

pytest -v -s python/fmha_sm100/cute/test_sparse_atten.py
pytest -v -s python/fmha_sm100/cute/test_fp4_indexer.py
pytest -v -s tests

E2E Minimax:

vllm serve nvidia/MiniMax-M3-NVFP4 \
  --tensor-parallel-size=8 \
  --attention_config.indexer_kv_dtype fp8 \
  --gpu-memory-utilization 0.95 \
  --max-model-len 2304 \
  --kv-cache-dtype fp8 \
  --block-size 128 \
  --language-model-only \
  --max-cudagraph-capture-size 2048 \
  --max-num-batched-tokens 2048 \
  --stream-interval 20 \
  --no-enable-prefix-caching \
  --trust-remote-code \
  --no-enable-flashinfer-autotune

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8992|±  |0.0083|
|     |       |strict-match    |     5|exact_match|↑  |0.8984|±  |0.0083|

[arpera]

Currently vllm requires quack-kernels>=0.4.0. In that case do we need to align vllm requirements with MSA requirements?

[lengrongfu]

I can in this pr vllm-project/vllm#49016 to together update.

[wzhao18]

@arpera Should we pin to an exact version? With quack-kernels>=xxx I believe it will just install the latest version.

@lengrongfu Sounds good. We can update the version there when we update the MSA commit hash. Thanks!

[arpera]

I think there is no need to pin an exact version here since we only need at least 0.6.1. Also, leave please a comment in the files where you update version of quack-kernels to know exactly why we need at least 0.6.1 installed.

[ZJY0516]

cc @yongye