sha2: fix AVX backend

[newpavlov]

Closes #344

[elichai]

This is much nicer!

[elichai]

On my machine(i9-9980HK) with turbo boost off, ran it a few times to make sure I get consistent result

Before:

test sha256_10    ... bench:          80 ns/iter (+/- 6) = 125 MB/s
test sha256_100   ... bench:         808 ns/iter (+/- 99) = 123 MB/s
test sha256_1000  ... bench:       7,734 ns/iter (+/- 402) = 129 MB/s
test sha256_10000 ... bench:      76,839 ns/iter (+/- 24,415) = 130 MB/s
test sha512_10    ... bench:          42 ns/iter (+/- 1) = 238 MB/s
test sha512_100   ... bench:         405 ns/iter (+/- 12) = 246 MB/s
test sha512_1000  ... bench:       3,532 ns/iter (+/- 587) = 283 MB/s
test sha512_10000 ... bench:      33,786 ns/iter (+/- 651) = 295 MB/s

After:

test sha256_10    ... bench:          80 ns/iter (+/- 1) = 125 MB/s
test sha256_100   ... bench:         789 ns/iter (+/- 11) = 126 MB/s
test sha256_1000  ... bench:       7,727 ns/iter (+/- 28) = 129 MB/s
test sha256_10000 ... bench:      76,364 ns/iter (+/- 1,379) = 130 MB/s
test sha512_10    ... bench:          40 ns/iter (+/- 0) = 250 MB/s
test sha512_100   ... bench:         390 ns/iter (+/- 22) = 256 MB/s
test sha512_1000  ... bench:       3,497 ns/iter (+/- 117) = 285 MB/s
test sha512_10000 ... bench:      33,845 ns/iter (+/- 7,469) = 295 MB/s

[elichai]

Just ran the same code in #344 and it doesn't crash

$ RUSTFLAGS='-Clink-arg=-fuse-ld=lld -Clinker-plugin-lto' cargo r
    Finished dev [unoptimized + debuginfo] target(s) in 0.01s
     Running `target/debug/wat`
[114, 100, 141, 235, 57, 20, 22, 206, 85, 130, 13, 65, 63, 33, 253, 46, 205, 24, 97, 124, 98, 149, 88, 88, 5, 158, 217, 104, 21, 56, 14, 150, 12, 43, 176, 5, 247, 183, 146, 200, 42, 155, 182, 2, 46, 191, 224, 193, 58, 249, 124, 43, 226, 194, 67, 166, 58, 22, 38, 122, 42, 128, 55, 218]