Poseidon evm hasher optimizations for lower input widths

[kilic]

With lower values of t there are some tricks that can be applied to reduce gas cost.

• Use add instead of addmod in mds multiplication and round constant phases. bn254 base field is suitable for deferred reductions upto 5 since 2**256 / Q ~= 5.29
• Use stack intensively by storing some of matrix cells at stack.

Applying these for (3, 8, 55) setup, I'm observing reduction of gas cost from 35k to 21k. Here you can check the example.