copies: M2MF from 16 KB up, SSE1 below; drop the toggle and stats

Removes the experiment scaffolding (glM2MCopyEnable/Enabled/GetStats, GLM2MCopyStats, the per-copy
TSC timing). gli_memcpy now takes the engine unconditionally at >= 16 KB and the SSE1 streaming path
otherwise; an unavailable engine or a failed copy falls through to SSE1, so correctness never depends
on M2MF. glMemcpyNV2A stays as the exported copy primitive.
