Auto-select sparse MLA single-token decode launch config

[xzyaoi]

Finding

The sparse MLA config module documents a measured single-token decode config that is 1.13-1.24x faster at Tq=1, but it is opt-in through XKERNELS_SPARSE_MLA_CONFIG:

• Measured result and default rationale: src/xkernels/ops/attention/triton/sparse_mla_config.py#L52-L64
• Opt-in DECODE_SPARSE_MLA_CONFIG: src/xkernels/ops/attention/triton/sparse_mla_config.py#L74-L85
• The launcher knows T before resolving config, but currently calls resolve_sparse_mla_config() without passing it: src/xkernels/ops/attention/triton/sparse_mla_kernel.py#L129-L160

Why this should improve performance

Single-token decode is a common serving hot path. The measured faster config is not hidden or speculative; it is already checked in, but callers have to know to export an environment override. Selecting it automatically when T == 1 would capture the win without regressing the documented multi-token case.

Suggested implementation

• Change config resolution to accept T or a decode_mode flag.
• If no env override is set and T == 1, return DECODE_SPARSE_MLA_CONFIG; otherwise return DEFAULT_SPARSE_MLA_CONFIG.
• Keep the env override as the highest-priority path for A/B testing and operator control.

Validation

• Unit-test config resolution for T=1, T>1, and env override cases.
• Re-run sparse MLA benchmarks for T=1, T=8, and T=64 to confirm the auto branch keeps the documented win and avoids multi-token regression.