Allow to specify floating point precision when fitting affine models

[igorpisarev]

This is a revised (less intrusive) version of #44. Added an overload for fit() to affine model classes that takes a user-specified epsilon for testing whether the configuration is ill-defined or not.