Remove cleverness in Bisection linesearch

[btracey]

The function comparison trickiness in Bisection causes difficulty with the operation of the rest of the package. Furthermore, it doesn't even work in the general case -- the floating point noise depends on the specific function. It's just not possible to converge a general non-linear function that far. Instead, we should at least describe (and possibly provide) a simple way to converge the gradient past 1e-10 if the user is willing to assume the function is locally quasi-convex.