Improve re_accurate

[RipleyTom]

• After some testing the full sequence on real ps3 it seems that frest => fi => fnms => fma sequence will differ from 1/x in 2 ways:
  Infinite on x86 becomes NaN(with sign respected) on ps3
  NaN on x86 becomes Zero(with no sign) on ps3
  Fixed the code to reflect it.
• Added the 'one' value as a parameter to the intrinsic(might make it easier to add intrinsic for other special values if any is found.)
• fixed 2 messed up patterns(checked a for both fnms and frest, oops).
• Removed 2 useless if checks that predates the xfloat merging PR from ages ago.
• on re_accurate, erased all the stores(might lead to some regressions, I tried a bunch of games and they all seemed fine).
• Fixed full fm(a, re_accurate) pattern with the same pattern as 1/x(as 0 * value => 0 and nan * value => nan).
• Since that fixes Resistance 2 bug with this pattern, put it back into approximate.

Fixes #15186