Approx rework: default scale == 0, epsilon applies to Approx::value

Also adds check to Approx::epsilon that the new epsilon has a valid
(ie one between 0 and 1)

Based on
http://realtimecollisiondetection.net/blog/?p=89
https://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/
https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html
https://en.wikipedia.org/wiki/Approximation_error#Formal_Definition

The given epsilon should refer to the target value, otherwise
the result would be unexpected, e.g. 101.02 == Approx(100).epsilon(0.01)
is true.
The default scale should be invisible, thus,
e.g. 101.01 == Approx(100).epsilon(0.01) gets false.
Finally even 101.000001 == Approx(100).epsilon(0.01) is false

include/internal/catch_approx.cpp

@@ -23,7 +23,7 @@ namespace Detail {
     Approx::Approx ( double value )
     :   m_epsilon( std::numeric_limits<float>::epsilon()*100 ),
         m_margin( 0.0 ),
-        m_scale( 1.0 ),
+        m_scale( 0.0 ),
         m_value( value )
     {}