nx_q4k_gemm_mt.nx

nx_q4k_gemm_mt.nx -- multi-threaded integer GEMM: wire nx_thread into the Q4_K integer dot. sd-server -> Nishi migration (cross-hw census PARTIAL -> HAVE: threading). The GEMM is embarrassingly parallel across output neurons: each worker thread computes a disjoint output range via nx_q4k_dot_row_col (its own iterator), joined by a done-flag (matches the nx_thread_spawn smoke: raw *i64 ctx, spin-join). Works on ANY multicore CPU. Verified bit-exact vs single-thread + measured speedup. ctx (flat *i64, 9 slots): [0]buf [1]w_off [2]o_start [3]o_end [4]n_blocks [5]rstride [6]col [7]out [8]done license_tier: ORIGINAL