Optimize CRC24 using pclmulqdq

[dignifiedquire]

There is a rust implementation for the non streaming version in https://github.com/TobiasBengtsson/crc24-openpgp-fast. It should be possible to implement a version that is streaming based on the paper and this code with the logic

• init state
• update when a chunk comes in
  • combine with existing stored state
  • if >= 2 * 4 * 128 bits, fold_by_4 reduction, store state, store rest
  • when it is less, just store the new rest
• finish
  • do the regular reduction based on the current state, the stored rest and fold_by_1 and barret reduction

This would speed up both reading and writing armored data