v2.0.4: Changes:

- Fixed inconsistencies in scalar_traits usage that yielded incorrect results for mixed integer/floating-point calculations.

- Added cml::epsilon<>, and changed most defaulted tolerance arguments to use it.