Improve FixedError root finding algorithm.

[carynbear]

Current Implementation:

• Exponential search forward until a bound is found where n_left is above and n_right is below target error.
• Binary search to find lowest n that is below target error.
• Assumes error is monotonically decreasing

Old Implementation of Root Finding with Backtracking, now deleted:

• Assumes error is convex
• Exponential search forward until a bound is found where n_left is above and n_right is below target error.
  • If we overshoot (error is increasing), backtrack two steps and restart exponential search with smaller increment
• Binary search to find lowest n that is below target error, assuming error is convex

In reality, test error is not smooth therefore we apply a smoothing window.