more robust CPU feature compilation and runtime selection

[hannesm]

at the moment (with #96 merged):

• __builtin_bswap is avoided (since together with -mssse3 it leads to SSSE3 instructions that won't work on e.g. AMD Phenom II)
• chacha source is copied to be compiled twice, once with and once without -mssse3

newer compilers with more analysis may optimize even more code into ssse3 instructions (of other hashes / ...).

to avoid this potential miscompilation, "all" C code should be compiled twice:

• once with machine intrinsics flags that are useful (atm ssse3/aes/pclmul)
• once without any machine-dependent intrinsics

then all entries (called from OCaml) should be runtime dispatched on the specific feature flags.

the issue with the above approach is that it is not yet clear which CPU features could be used in which settings and which features are useful (thinking of SSSE4 etc. as well). to avoid a huge matrix, time should be used to research what is useful.