CRC32/MD5 optimisations

[animetosho]

I noticed that a few simple things could be done to improve hashing performance:

• CRC32 zero extension can be computed using GF arithmetic instead of passing 0s through. Doing so greatly speeds up the window table generation
• the CRC window mask can be incorporated into the window table itself, saving the mask operation. Speed difference is likely negligible, but it simplifies code
• avoid unnecessarily copying blocks in MD5 on little endian CPUs
• slightly faster MD5 round expressions

(I don't really see par2cmdline being highly performance focused, but these are relatively simple tweaks which I thought would make sense to have)