crc32: align memory access

There is some assembly working with a byte array like with an
array of unsigned long values. That is incorrect, because the byte
array may be not aligned by 'unsigned long'.

The patch makes crc32 calculate the hash on the prefix
byte-by-byte, then word-by-word on aligned addresses, and again
byte-by-byte on a tail which is less than word.

Assuming that the word-by-word part is the longest, this should
reduce number of memory/cache loads in ~x2 times. Because in case
of a not aligned word load it was necessary to load 2 words, and
then merge them into one. When addresses are aligned, this is only
1 load.

Part of #4609