More lz_hash_chains, lz_binary_trees performance improvements

- Use a multiplicative hash function.  In practice it's as good as the
  CRC32-based hash function previously used, but it's faster to compute
  and requires no static data.

- Use slightly different logic that avoids the need to special-case the
  extension of matches from 'nice_len' to 'max_len'.

- Faster skip_positions() by avoiding unnecessary calculations.

- Fast match length extension on x86 using unaligned word accesses and
  trailing zero count.