[None][feat] VisualGen VSA and SOL sparse attention on PrimTS block-sparse FMHA (overview)

[heyuhhh]

Description

Overview PR for VisualGen VSA and SOL sparse attention on generic PrimTS block-sparse FMHA, kept for end-to-end
review of the full stack. The mergeable units are the split PRs below; this branch carries all six commits so the
complete change can be read and tested in one place.

Review stack

1. Generic PrimTS block-sparse FMHA and sparse-attention framework refactor (4 commits, targets main), including
   the core TrtllmAttention.block_sparse_attn_predict hook, the SparseBackendForwardArgs.block_sparse_inputs
   carrier that algorithms use to hand routes to the core, prepare_sparse_runtime_params as the single aggregation
   point (all prediction hooks plus SkipSoftmax scheduling), and the FA4 WorkTileInfo compatibility fix:
   NVIDIA/TensorRT-LLM#18815.
2. VisualGen VSA (predicts before the core forward and hands routes through sparse_backend_args) and SOL (overrides
   the core hook) algorithms (2 commits, stacked on the branch above): [None][feat] add VisualGen VSA and SOL sparse attention on PrimTS block-sparse FMHA heyuhhh/TensorRT-LLM#10.
3. FlashInfer kernels: proxy-compensated block-sparse attention feat(prims-ts): support proxy-compensated block-sparse attention flashinfer-ai/flashinfer#4872 (merged) and PrimTS decode
   optimizations perf(prims-ts): Optimize&refine PrimsTS block sparse attention flashinfer-ai/flashinfer#5002 (pending); the vendored PrimTS tree is pinned to
   heyuhhh/flashinfer@61454c5c (yuhangh/tmp-sol-attn-trtllm-dev), which is yuxianq/flashinfer:trtllm-prims-ts-dev
   (the tree main pins) plus [TRTLLM-5589] feat: Integrate TRT-LLM Gen FP8 Batched GEMM with Pytorch workflow kernel autotuner #4872, Feature request: pin versions of python packages in CI #5002 and one follow-up commit (fixed 2D block_tables for both paged
   block-sparse entry points, a dense-style validate switch on the wrappers, and the PrimTS block-sparse trace
   templates rewritten as literal schemas).

Commits

1. chore: update FlashInfer PrimTS pin
2. feat: add generic PrimTS block-sparse support
3. refactor: unify sparse attention runtime inputs
4. fix: keep FA4 WorkTileInfo independent of CUTLASS task scheduling
5. refactor: align VisualGen sparse attention workflow
6. feat: add VisualGen SOL sparse attention

Test Coverage

See the split PRs for the per-layer test runs.