Add reusable workspaces for hot decode kernels and graph capture

[xzyaoi]

Finding

Several hot-path wrappers allocate and zero temporary tensors on every call:

• MoE align allocates sorted_ids, expert_ids, num_post, tokens_cnts, and cumsum every call: src/xkernels/ops/moe/triton/align_kernel.py#L189-L193
• INT4 fused MoE allocates either [M, N] fp32 combine output or [M * top_k, N] scratch: src/xkernels/ops/moe/triton/moe_int4_kernel.py#L441-L464
• MXFP4 MoE allocates the activation scratch and zeroed combine output: src/xkernels/ops/moe/triton/moe_mxfp4_kernel.py#L436-L452
• Sparse MLA attention allocates out, lse, and maxl every call: src/xkernels/ops/attention/triton/sparse_mla_kernel.py#L134-L136

Why this should improve performance

In decode serving, many of these kernels run once per layer per token. Per-call allocation and memzero overhead can dominate small-batch latency and complicates HIP/CUDA graph capture. The code already has some fixed-shape modes (truncate=False in MoE align), so the next step is to let callers reuse the fixed buffers.

Suggested implementation

• Add optional workspace= / out= arguments for the hot public APIs, or small workspace dataclasses per op family.
• Size workspaces from existing deterministic bounds (max_pad, max_blocks, M, top_k, N, hidden size).
• Keep allocating by default for ergonomic use, but let serving stacks preallocate once per max decode shape.
• Avoid zeroing buffers when the kernel overwrites all live elements; keep explicit zeroing only for atomic-add combine outputs or EP partial outputs that need zero for skipped experts.

Validation

• Add tests that pass preallocated buffers and compare results with allocation-owning calls.
• Benchmark small decode buckets with allocation-owning eager, preallocated eager, and graph-captured execution.
• Include a stress test that reuses a larger workspace for smaller M buckets without stale-data leakage.