chacha20: Improve AVX2 performance

[str4d]

• A bunch of instructions for accessing the 128-bit lanes have been replaced by a union.
• I've implemented a widely-used ChaCha optimisation (that I spotted in c2-chacha).

[str4d]

Ran cargo +nightly bench -p chacha20 in WSL2 on my desktop (i7-8700K overclocked to 4.8GHz).

$ cargo +nightly --version
cargo 1.55.0-nightly (d21c22870 2021-07-26)

Current master:

     Running unittests (target/release/deps/chacha12-f20b37e08443492e)
test bench1_10     ... bench:           9 ns/iter (+/- 0) = 1111 MB/s
test bench2_100    ... bench:          52 ns/iter (+/- 1) = 1923 MB/s
test bench3_1000   ... bench:         739 ns/iter (+/- 53) = 1353 MB/s
test bench4_10000  ... bench:       7,576 ns/iter (+/- 277) = 1319 MB/s
test bench5_100000 ... bench:      76,180 ns/iter (+/- 3,424) = 1312 MB/s

     Running unittests (target/release/deps/chacha20-5391fc4cd79914d1)
test bench1_10     ... bench:          12 ns/iter (+/- 0) = 833 MB/s
test bench2_100    ... bench:          80 ns/iter (+/- 7) = 1250 MB/s
test bench3_1000   ... bench:         952 ns/iter (+/- 26) = 1050 MB/s
test bench4_10000  ... bench:       9,765 ns/iter (+/- 407) = 1024 MB/s
test bench5_100000 ... bench:      97,608 ns/iter (+/- 3,413) = 1024 MB/s

     Running unittests (target/release/deps/chacha8-58d30d30e94644a2)
test bench1_10     ... bench:           8 ns/iter (+/- 0) = 1250 MB/s
test bench2_100    ... bench:          41 ns/iter (+/- 2) = 2439 MB/s
test bench3_1000   ... bench:         625 ns/iter (+/- 21) = 1600 MB/s
test bench4_10000  ... bench:       6,495 ns/iter (+/- 217) = 1539 MB/s
test bench5_100000 ... bench:      65,487 ns/iter (+/- 3,186) = 1527 MB/s

After introducing backend::avx2::StateWord union:

     Running unittests (target/release/deps/chacha12-f20b37e08443492e)
test bench1_10     ... bench:          10 ns/iter (+/- 0) = 1000 MB/s
test bench2_100    ... bench:          50 ns/iter (+/- 1) = 2000 MB/s
test bench3_1000   ... bench:         411 ns/iter (+/- 22) = 2433 MB/s
test bench4_10000  ... bench:       4,053 ns/iter (+/- 207) = 2467 MB/s
test bench5_100000 ... bench:      39,887 ns/iter (+/- 1,633) = 2507 MB/s

     Running unittests (target/release/deps/chacha20-5391fc4cd79914d1)
test bench1_10     ... bench:          12 ns/iter (+/- 0) = 833 MB/s
test bench2_100    ... bench:          74 ns/iter (+/- 1) = 1351 MB/s
test bench3_1000   ... bench:         636 ns/iter (+/- 38) = 1572 MB/s
test bench4_10000  ... bench:       6,189 ns/iter (+/- 690) = 1615 MB/s
test bench5_100000 ... bench:      61,537 ns/iter (+/- 2,681) = 1625 MB/s

     Running unittests (target/release/deps/chacha8-58d30d30e94644a2)
test bench1_10     ... bench:           8 ns/iter (+/- 0) = 1250 MB/s
test bench2_100    ... bench:          38 ns/iter (+/- 2) = 2631 MB/s
test bench3_1000   ... bench:         303 ns/iter (+/- 12) = 3300 MB/s
test bench4_10000  ... bench:       2,910 ns/iter (+/- 98) = 3436 MB/s
test bench5_100000 ... bench:      28,850 ns/iter (+/- 1,079) = 3466 MB/s

After diagonalization optimization:

     Running unittests (target/release/deps/chacha12-a4a957c0b61fe24c)
test bench1_10     ... bench:          10 ns/iter (+/- 0) = 1000 MB/s
test bench2_100    ... bench:          49 ns/iter (+/- 3) = 2040 MB/s
test bench3_1000   ... bench:         391 ns/iter (+/- 14) = 2557 MB/s
test bench4_10000  ... bench:       3,799 ns/iter (+/- 95) = 2632 MB/s
test bench5_100000 ... bench:      37,872 ns/iter (+/- 1,195) = 2640 MB/s

     Running unittests (target/release/deps/chacha20-cb419216b50db2c1)
test bench1_10     ... bench:          12 ns/iter (+/- 0) = 833 MB/s
test bench2_100    ... bench:          72 ns/iter (+/- 3) = 1388 MB/s
test bench3_1000   ... bench:         599 ns/iter (+/- 21) = 1669 MB/s
test bench4_10000  ... bench:       5,858 ns/iter (+/- 308) = 1707 MB/s
test bench5_100000 ... bench:      58,155 ns/iter (+/- 2,432) = 1719 MB/s

     Running unittests (target/release/deps/chacha8-276d7f24b799657c)
test bench1_10     ... bench:           8 ns/iter (+/- 0) = 1250 MB/s
test bench2_100    ... bench:          37 ns/iter (+/- 1) = 2702 MB/s
test bench3_1000   ... bench:         291 ns/iter (+/- 12) = 3436 MB/s
test bench4_10000  ... bench:       2,792 ns/iter (+/- 76) = 3581 MB/s
test bench5_100000 ... bench:      27,780 ns/iter (+/- 1,477) = 3599 MB/s