perf: Use explicit squaring chain for the x^5 S-box

[vadorovsky]

Field::pow is a generic binary exponentiation, costing 3 squarings + 2 multiplications for x^5. Replacing it with x^2 -> x^4 -> x^4 * x (2 squarings + 1 multiplication) speeds up hashing by up to ~47% for 1-2 inputs, degrading to ~2% at 12 inputs where the MDS dominates.

Benchmarked on AMD EPYC 7H12 64-core with the following results.

Before:

poseidon_bn254_x5_1     time:   [22.458 µs 22.501 µs 22.554 µs]
poseidon_bn254_x5_2     time:   [33.189 µs 33.391 µs 33.602 µs]
poseidon_bn254_x5_3     time:   [44.983 µs 45.054 µs 45.130 µs]
poseidon_bn254_x5_4     time:   [63.479 µs 63.533 µs 63.596 µs]
poseidon_bn254_x5_5     time:   [83.178 µs 83.214 µs 83.247 µs]
poseidon_bn254_x5_6     time:   [112.51 µs 112.81 µs 113.17 µs]
poseidon_bn254_x5_7     time:   [142.82 µs 143.19 µs 143.65 µs]
poseidon_bn254_x5_8     time:   [175.62 µs 176.55 µs 177.72 µs]
poseidon_bn254_x5_9     time:   [203.31 µs 203.36 µs 203.42 µs]
poseidon_bn254_x5_10    time:   [275.25 µs 276.37 µs 277.65 µs]
poseidon_bn254_x5_11    time:   [296.42 µs 296.77 µs 297.26 µs]
poseidon_bn254_x5_12    time:   [374.93 µs 375.45 µs 376.07 µs]

After:

poseidon_bn254_x5_1     time:   [15.333 µs 15.348 µs 15.367 µs]
poseidon_bn254_x5_2     time:   [24.189 µs 24.213 µs 24.244 µs]
poseidon_bn254_x5_3     time:   [35.364 µs 35.392 µs 35.427 µs]
poseidon_bn254_x5_4     time:   [53.129 µs 53.360 µs 53.639 µs]
poseidon_bn254_x5_5     time:   [71.739 µs 71.854 µs 71.995 µs]
poseidon_bn254_x5_6     time:   [98.231 µs 98.314 µs 98.407 µs]
poseidon_bn254_x5_7     time:   [128.92 µs 129.67 µs 130.74 µs]
poseidon_bn254_x5_8     time:   [163.55 µs 164.76 µs 166.04 µs]
poseidon_bn254_x5_9     time:   [190.60 µs 191.32 µs 192.08 µs]
poseidon_bn254_x5_10    time:   [256.79 µs 257.19 µs 257.65 µs]
poseidon_bn254_x5_11    time:   [282.91 µs 283.55 µs 284.33 µs]
poseidon_bn254_x5_12    time:   [367.30 µs 369.78 µs 372.52 µs]

inputs	baseline	sbox opt	speedup
1	22.50 µs	15.35 µs	1.47×
2	33.39 µs	24.21 µs	1.38×
3	45.05 µs	35.39 µs	1.27×
4	63.53 µs	53.36 µs	1.19×
5	83.21 µs	71.85 µs	1.16×
6	112.81 µs	98.31 µs	1.15×
7	143.19 µs	129.67 µs	1.10×
8	176.55 µs	164.76 µs	1.07×
9	203.36 µs	191.32 µs	1.06×
10	276.37 µs	257.19 µs	1.07×
11	296.77 µs	283.55 µs	1.05×
12	375.45 µs	369.78 µs	1.02×

[vadorovsky]

I know that the same trick is done in #54, but I think it's better to keep one scoped optimization per PR instead of mixing them together. I'm planning to address performance of state, MDS etc. in a bit different way, and also get rid of vectors in all possible places.