Thanks Johannes for the review.

This patch hopefully fixes the problems you mentioned:

- Hexadecimals, binary and numbers with `_` are considered a single
  token
	Pre                    Post
    0xFF_EC_DE_5E          0xFF_E1_DE_5E
	0b100_000              0b100_100
	100_000                200_000

  Even though a single character is changed in each of the above
  numbers, the diffs would be produced as if they were single tokens


- More tests added for "proper" multicharacter operators.
  
  Earlier regex would consider a++!=++b as 3 different tokens(a, ++!=++, b)
  This patch matches the tokens properly into (a, ++, !=, ++, b)