Remove permute() and alternate method of getting g() inlined leads to better performance

[renthraysk]

• Factors out the need for permute(), by generating code for each round.
• g() inlines when pass the values instead of an array pointer and indices. This allows putting state across 16 variables instead of [16]uint32 which then causes the compiler to use more registers.

name      old time/op    new time/op    delta
Write-4     6.18ns ± 0%    3.41ns ± 0%  -44.93%  (p=0.000 n=9+10)
Sum256-4    6.21µs ± 0%    3.51µs ± 0%  -43.43%  (p=0.000 n=9+9)
XOF-4       5.73ns ± 0%    3.13ns ± 0%  -45.36%  (p=0.000 n=9+10)

name      old speed      new speed      delta
Write-4    162MB/s ± 0%   294MB/s ± 0%  +81.53%  (p=0.000 n=9+10)
XOF-4      174MB/s ± 0%   319MB/s ± 0%  +83.04%  (p=0.000 n=9+10)

name      old alloc/op   new alloc/op   delta
Write-4      0.00B          0.00B          ~     (all equal)
Sum256-4     0.00B          0.00B          ~     (all equal)
XOF-4        0.00B          0.00B          ~     (all equal)

name      old allocs/op  new allocs/op  delta
Write-4       0.00           0.00          ~     (all equal)
Sum256-4      0.00           0.00          ~     (all equal)
XOF-4         0.00           0.00          ~     (all equal)

[lukechampine]

Thanks again -- another huge increase!

I decided not to keep the codegen, since it's not likely to ever change, and the generated code isn't terribly hard to verify. However, if someone wants to add an asm implementation, I would push for codegen there -- a tool like avo would make the asm much easier to to review and verify.