[AutoDeploy]: MLA optimizations for DS-R1

[nzmora-nvidia]

🚀 The feature, motivation and pitch

• Torch operators: MLA MHA-mode (no weight absorption) with and without cache.
• FlashInfer operators: MLA MQA-mode (weight absorption) with cache; for decode and mixed decode+prefill (flattened). Uses flashinfer.mla.BatchMLAPagedAttentionWrapper and flashinfer.append_paged_mla_kv_cache
• Weight absorption optimizer pass
• Patch deepseek R1
• Pytorch and FlashInfer MLA backends
• Cuda graph for FlashInfer MLA operator
• All correctness tests are passing

In a future task:

• FlashInfer MLA MHA-mode (no weight absorption) with cache; ragged; for prefill-only. Uses flashinfer.BatchPrefillWithRaggedKVCacheWrapper
  This kernel is performant in prefill-only use-cases.
  To support mixed decode+prefill we need to:
  1. Compute new ckv + k_pe and append to the cache (paged)
  2. Read from the cache and write to a new ragged layout (paged cache has "holes") and the kernel only uses kv_indptr without lengths.
  3. Compute the output

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[galagam]

Draft PR
Rebasing @nzmora-nvidia's code on top of main tot + fixing some conflicts + missing APIs.
Need to test and understand better before submitting for review.

#10504