keccak: remove the portable SIMD backend?

[newpavlov]

The portable SIMD implementation requires a Nightly compiler and could be useful only for algorithms which are able to use parallel processing. Additionally, users have to specify desired SIMD width manually, which is pretty inconvenient and likely to be done extremely rarely.

Thus I think it may be worth to remove the portable SIMD support completely. At the very least, until portable SIMD is stabilized and there will be a clear picture on how to select optimal SIMD width at runtime.

cc @aewag

[tarcieri]

I think the relevant algorithms can all make use of as much SIMD width as is available

[newpavlov]

AFAIK the big problem is that portable SIMD does not provide a way to find how much "SIMD width is available". It also does not handle enabling of necessary target features. For example, even if we wrote our code over u64x4 we still need to manually detect AVX-512 and use target_feature(enable="avx512...") on relevant functions, otherwise the code will be compiled using SSE instructions, likely with a lot of stack spilling.