2.0.19

Fixes

• OpenCL: fix ~2x fp32 throughput undercount on Apple. The compute
  kernels used the mad() builtin, which the OpenCL spec permits (but does
  not require) to fuse into hardware FMA. Apple's deprecated CL-on-Metal 1.2
  compiler never lowers it, silently halving measured fp32 throughput
  (2.2 → 4.5 TFLOPS on M1 Pro). Rewritten as a plain contracted expression
  ((y*x)+y), which forces the compiler to choose between fusing or
  violating -cl-mad-enable — all vendors honor this. OpenCL fp32 now
  matches Metal on Apple (within 1–2%); other platforms are unaffected
  (mad() already fused there).
• CUDA/ROCm: mark global-bandwidth kernel pointers __restrict__. The
  in/out pointer pair had no aliasing annotation, so the compiler
  couldn't batch loads ahead of the trailing store and NVIDIA couldn't take
  the read-only ld.global.nc path. These were the only kernels in either
  backend with two pointer parameters, so the fix is contained to them.
• CPU: fix GCC build failure on the crypto ISA extensions (#189).
  GCC declares AES/SHA-256 NEON intrinsics under a single umbrella
  +crypto target attribute, so a translation unit built with only the
  granular +aes or +sha2 flag failed to inline them ("target specific
  option mismatch"). GCC now gets +crypto (+crypto+sha3 for the SHA-512
  TU); Clang keeps the granular flags, since +crypto is deprecated there.

CI

• Added a GCC-only build job to catch this class of compiler-specific
  regression going forward.

Full Changelog: 2.0.18...2.0.19