add runtime feature detection for avx2

[alindima]

Currently, the avx feature is compile-time determined, based on the CPUID features of the processor and/or RUSTFLAGS value.

Before cutting a new release that includes #10, we must add runtime feature detection.
Otherwise, one could use pre-compiled binaries from a platform that supports avx and would likely run into an illegal instruction when running on a platform that doesn't.

[ordian]

Take a look at dalek-cryptography/curve25519-dalek#523 for inspiration.