Various cleanups, including tweaks to make the Adler-32 code more
consistent with the CRC-32 code and vice versa.  No behavior changes.

- As was recently done in the Adler-32 code, take advantage of the fact
  that on recent x86 processors, vmovdqu with an aligned pointer is just
  as fast as vmovdqa.  Don't waste time aligning the pointer unless the
  length is very large, and at the same time, handle all cases of
  len >= 8*VL using the main loop so that the 4*VL wide loop isn't
  needed.  (Before, aligning the pointer was tied to whether the main
  loop was used or not, since the main loop used vmovdqa.)

- Handle short lengths more efficiently.  Instead of falling back to
  crc32_slice1() for all len < VL, use AVX-512 masking (when available)
  to handle 4 <= len <= 15, and use 128-bit vector instructions to
  handle 16 <= len < VL.

- Document why the main loop uses a width of 8*VL instead of 4*VL.