Find a balance between head only reduction and full reduction.

[lvella]

Performing head only reduction improved performance for the vast majority of test cases, still some suffered a minor performance hit. Indeed, it is possible that fully reducing just the hottest reducers (the base elements most used in reductions) will yield a great performance gain.

For one case, I found that 13% of the polynomials are responsible for 95% of reductions performed, so just fully reducing a polynomial when it gets used in more than 4 or 5% of past reductions might yield a net gain in performance.

[lvella]

I have investigated a bit into this in the lazy_reduction branch, but never managed to improve over only reducing the leading term.