Switch from using floats to fixed point backed by integers

There is some evidence that fixed point arithmetic should be faster than float32.

Some libraries:

• https://github.com/PetteriAimonen/libfixmath
• https://github.com/PetteriAimonen/libfixmatrix
• https://github.com/chmike/fpsqrt
• https://github.com/jcarrano/fixed_point_arith