Improving speed/quality

[Bulat-Ziganshin]

I have two suggestions for this code:

foldhash/src/lib.rs

Line 314 in 8f878c6

s0 = folded_multiply(load(v, 0) ^ s0, load(v, 32) ^ seeds[0]);

1. use more variables. MUL latency is 3 (and 4 for higher 64 bits afaik), other operations 1, so at least 6 vars required to hide latency
2. mix vars. Instead of seed[0], use next var - i.e. s1 for s0 computation and so on. This way, entropy from a single input bit will spread around all vars. It will allow one to use it as a proper 128-256 bit hash if she needs one