Improve axis regular precision

[ryanelandt]

This PR refactors regular axis to address the issues described in #336 by incorporating some of the ideas from #386. This PR adds a unit that verifies exact indexing for various cases.

The output of ./benchmark_axis_index before the refactor:

Benchmark                         Time             CPU   Iterations
-------------------------------------------------------------------
regular<uniform>               1.31 ns         1.31 ns    538240574
regular<normal>                2.00 ns         2.00 ns    351784388
circular<uniform>              1.62 ns         1.62 ns    432490637
circular<normal>               1.62 ns         1.62 ns    424151491

The output of ./benchmark_axis_index after the refactor:

Benchmark                         Time             CPU   Iterations
-------------------------------------------------------------------
regular<uniform>               1.13 ns         1.13 ns    578725526
regular<normal>                1.71 ns         1.71 ns    399524265
circular<uniform>              1.51 ns         1.51 ns    471027721
circular<normal>               1.49 ns         1.49 ns    459937000

[ryanelandt]

This PR does not change the public API, but it does change private class members that are not user accessible. Modifying private class members has implications for serialization. Is this type of PR problematic in regards to Boost API stability?

[emmenlau]

This looks actually quite promising, and performance is better than with our (naiive) solution. Could this be considered for integration?

[henryiii]

By the way, I think this would help scikit-hep/hist#455.