Deprecated --auto-detect option. By default, the application
will now test all interfaces supported by the processor.

When cipher offset passed through job API is multiple of 8 bits,
instead of using less performant functions which handle offsets
in bits, the source and destination pointers can be adjusted
and more performant and simpler functions which cannot handle
offsets in bits are called.

- New API to select most performant architecture supported by the processor
- Added new architecture Enums to public header

Example of usage (run benchmark with packet sizes of 84, 256 and 1280 bytes,
with probabilities of 10%, 15% and 75%):

./perf/ipsec_perf --cipher-algo aes-docsis-128 --hash-algo docsis-crc32 \
                  --job-size 84,256,1280 --imix 10,15,75 --job-iter 3000000 \
                  -o 0 --arch AVX512

Job sizes can be passed now as a list of sizes and not just a single value
or a range.

Example (run benchmark with buffer sizes of 84B, 256B and 1280B):

./perf/ipsec_perf --cipher-algo aes-docsis-128 --hash-algo docsis-crc32 \
                  --job-size 84,256,1280 --job-iter 3000000 \
                  -o 0 --arch AVX512

- k-registers used to figure CRC status of the lane
- using masked operations (k-registers) to reduce amount of checks / branches in first and next/update cases
- first and update/next cases done always in parallel (ZMM width)
- last partial block handling is still sequential

avx512-vaes: eliminate store to load issue when loading plain text for CRC partial block (less than 12)

avx512-vaes: eliminated double store to load problem occuring in partial CRC blocks of size 12 to 15

This commit is part of fix for #62.

Co-authored-by: Martin Filla <martinfilla@sysctl.cz>

Array of pointers to keystream is copied to the stack
with aligned moves, so this array must be aligned to
16, 32 o 64 bytes, depending on the architecture.