Fix buffer overrun on a partial block in the C reference functions

[riha-dev]

encode_seq, encode_par, decode_seq and decode_par only compare len against zero while stepping it by the block size, so a length that is not a whole number of blocks wraps size_t and the loop runs off both buffers (adding a 1- or 2-byte case to the encode_exact table is enough to segfault the test binary); each now rounds len down to whole blocks first, which leaves the inner loop and its generated code untouched.

[ia0]

Thanks for the PR but the cmp crate is only used for benchmarks where partial blocks are uninteresting. The benchmarks only use full blocks, so there is no buffer overrun.