Generate Float and Double via division

[curiousleo]

Coverage

Before:

0.787% of representable Floats in the unit interval reached

After:

7.874% of representable Floats in the unit interval reached

(A similar enumeration for Double is impossible, but it is very likely
that coverage is increased for Doubles too.)

For the full analysis, see #105.

Performance

Before:

pure/random/Float                        mean 331.1 μs  ( +- 21.67 μs  )
pure/uniformR/unbounded/Float            mean 324.6 μs  ( +- 2.849 μs  )

pure/random/Double                       mean 411.3 μs  ( +- 5.876 μs  )
pure/uniformR/unbounded/Double           mean 416.8 μs  ( +- 41.93 μs  )

After:

pure/random/Float                        mean 27.32 μs  ( +- 158.0 ns  )
pure/uniformR/unbounded/Float            mean 27.37 μs  ( +- 422.0 ns  )

pure/random/Double                       mean 27.34 μs  ( +- 303.1 ns  )
pure/uniformR/unbounded/Double           mean 27.49 μs  ( +- 983.7 ns  )

---

This is simply a Pareto-improvement over what we currently have. #102 is related but separate, since it changes the coverage completely but at the cost of decreased performance.