sha256: improve hashing performance by starting w computation earlier

[Rexicon226]

While poking around the sha256 impl, I noticed some bad codegen around the schedule updates. GCC 15 and LLVM 21 seem to miss re-ordering the schedule updates higher for Zen 5 (LLVM still just aliases the numbers of Zen 4), and thus we miss on a bit of ILP.
With the benchmark provided in test_sha256.c, this increases the performance by around 1m hashes / second on a Ryzen 5 9600X.

LLVM 21.1.8:

NOTICE  01-02 02:07:37.666619 114269 f0   0    src/ballet/sha256/test_sha256.c(175): ~39.028 M poh hashes / sec / core with fd_sha256_hash_32_repeated
NOTICE  01-02 02:07:37.692786 114269 f0   0    src/ballet/sha256/test_sha256.c(185): ~38.221 M poh hashes / sec / core with fd_sha256_hash_32_repeated_old

GCC 15.2:

NOTICE  01-02 14:21:55.726875 679095 f0   0    src/ballet/sha256/test_sha256.c(175): ~39.057 M poh hashes / sec / core with fd_sha256_hash_32_repeated
NOTICE  01-02 14:21:55.753294 679095 f0   0    src/ballet/sha256/test_sha256.c(185): ~37.861 M poh hashes / sec / core with fd_sha256_hash_32_repeated_old

(also adds the missing FOUR_ROUNDS undefine, it was accidentally removed cuz it was a typo)