SHA-2 performance: keep the message schedule in locals

Rewrite the block functions of SHA224Digest, SHA256Digest and LongDigest
(SHA-384/512/512t) to hold the 16 schedule words in rotating locals and
run 16 unrolled rounds per pass, instead of expanding the full 64/80-word
schedule into an array first. The X/W buffers shrink to 16 words, which
also makes the reset() clear and Memoable copy 4-5x smaller.