Remove default `epsilon` implementation for `Float`

[AtheMathmo]

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.

[cuviper]

This issue was moved to rust-num/num-traits#4