gcm128.c: P.-M. Hager has tipped about possibility to fold reductions

in gcm_ghash_4bit. Taking the idea a step further I've added extra
256+16 bytes of per-key storage, so that one can speak about 3rd variant
in addition to "256B" and "4KB": "528B" one. Commonly it should be
~50% faster than "256B" implementation or ~25% slower than "4KB" one.