Remove default `epsilon` implementation for `Float`

[cuviper]

From @AtheMathmo on September 22, 2016 19:57

With #231 we introduced a new epsilon() function which returns the EPSILON constant for f32 and f64. In order to prevent breaking the API we provided a default implementation here - using NumCast to return f32::EPSILON.

When we are ready to release breaking changes we should consider removing this default implementation.

Copied from original issue: rust-num/num#232

[cuviper]

I've decided to limit the clutter of breaking change issues, so they will now be tracked collectively in #47. I'm closing this issue for now as postponed but we can still discuss its merits here as needed.