fixes some potential issues with underscores in float literals.
adds more checks for badly positionned underscores in float literals.
adds more test files.
Float literals were not parsed properly when their fractional part
exceeded 53 significant bits. This affected in particular math.PI
and math.E. Rather than reinventing the wheel, this patch reuses
C's strtod() implementation, which already does the heavy lifting
with respect to correctness, though some caution is necessary to
keep float parsing locale-independent.
