The flatness check in HagerZhang was intended to prevent excessive
iteration in cases where the objective function is flat to within
numerical precision. This test is a poor attempt at capturing the issue.
HZ takes more iterations than any other linesearch algorithm, but it's
still a far cry from real-world cases I've seen where the linesearch
exits from the iteration limit.

This may cause problems of its own, but for the time being it's better
than the status quo.

Fixes #173
Closes #174
Fixes #175

This follows the notation in the paper and is much more readable. It
also introduces an external parameter, `epsilonk`, that "solves" the
flatness problem (though it requires external specification).