Manually unroll MDS calculation

[georgeee]

Modify full_round to perform round inline when params.mds.len() == 3.

This optimization gave a ~3% speedup (measured by mina ledger test apply on top of a PR stack #2394, MinaProtocol/mina#15980 and TODO open a PR.

Although 3% isn't much, I expect speed-up to be better when a batch version of hash (MinaProtocol/mina#16053) is delivered.

P.S. change's diff is bloated by ^M characters present at the end of each line for the previous version of the poseidon/src/permutation.rs file.

[dannywillems]

As discussed with @georgeee online, we can close this for now.

[georgeee]

Copying my message from Slack

let's close this one for now. Initially I thought it would be useful. But now I'm not that sure. For the work I was doing I developed a quite different version of this computation (relying on parallelism). So the usage against which I was seeing 3% speed-up (i.e. ocaml bindings) will end up not involving the optimized call. And I have no idea whether this optimization would still make sense in other contexts (like WASM) or is likely to worsen things.
3% is also not that convincing. I thought there will be a follow-up with more increases, but it turned out to be a futile endeavor.