chacha20: Add wide (4-block) AVX2 impl

[str4d]

I've been digging into c2-chacha to figure out why its performance is consistently better than chacha20. I think I understand the main difference now: c2-chacha has a "wide" mode where it processes four ChaCha blocks at a time. This requires two sets of registers per state word (since a single 256-bit register only has room for two blocks in parallel), but I guess on modern CPUs there are enough registers that it can handle the increased number of temporaries, and the resulting interleaved instructions seem to parallelise well.