Add measured LLM decode kernel suite

[WaffleBits]

Summary

• add QK dot, paged KV gather, fused selected attention, residual RMSNorm, and signed INT4 GEMV Triton kernels
• check RTX 5070 Ti timing artifacts and a specification-based roofline report
• install a filtered Nsight Compute capture path with an explicit host counter-permission error

Verification

• ruff check .
• pytest -q (13 passed)
• measured GPU runs with correctness checks before timing

Claim boundary

Bandwidth and roofline traffic remain logical models. Nsight hardware counters are not claimed because Windows host counter access is disabled.