blake2b: Add aarch64 NEON acceleration

[r1mikey]

Import NEON acceleration from blake2b-ref. Copyrights reflect those in the upstream files, specifically:
https://github.com/BLAKE2/BLAKE2/blob/master/neon/blake2b-neon.c

I've restricted this to aarch64 as I do not have an environment in which to test armv7-a support.

Successfully built and run on both aarch64 (Nvidia Grace) and x86_64. Tested on both with make verify.

[r1mikey]

Let's hold on merging this while I do some benchmarking as well.

[jedisct1]

Something like this would be required after including <arm_neon.h>:

#    ifdef __clang__
#        pragma clang attribute push(__attribute__((target("neon"))), apply_to = function)
#    elif defined(__GNUC__)
#        pragma GCC target("+simd")
#    endif

[r1mikey]

This is not worth pursuing for 64 bit Arm systems based on my benchmarks. Performance of the NEON implementation comes in at around 66% of the C implementation. There are a few reasons for this:

• Instruction density is 89% of the C implementation (good)
• Instructions per cycle is at 46% of the C implementation (very bad)
• 54.6% of slots are backend bound (vs 21.3% for the C implementation - also very bad)
• The benchmark completes in 10,718.84 msec compared to 5,542.82 msec for the C implementation

[jedisct1]

Thanks for having taken the time to run these benchmarks!