main
Mali/UMA portability: fix non-coherent transfers + expand test suite 52->82 (#1)
Batch API: many dispatches, one submit + Fence drain-wait guard
Program & Launch modules: clspv reflection, compute pipelines, CUDA-style dispatch
Core modules: Context (volk init, device selection policy, compute queue, one-shot cmdbufs), VMA-backed Buffer with memcpy/staging transfer paths, Fence sync; exception-based XP_CHECK error handling