onnxruntime

5e2a1b55 - [CUDA] Add persistent fpA_intB MatMulNBits tactic autotune cache

[CUDA] Add persistent fpA_intB MatMulNBits tactic autotune cache Adds an opt-in, hardware-keyed on-disk tactic cache for the fpA_intB MatMulNBits path, a reduced first-time M-bucket sweep with lazy single-bucket fallback, CUDA EP teardown flush of lazily profiled tactics, an offline tuning tool, docs, and unit tests. Squashed and rebased onto origin/main.