ENH: handle underflow (rounding) of float/double as zero (issue #625)

- The problem occurs when using atof to parse values such as "1e-39"
  since this is out of range for a float and _can_ set errno to
  ERANGE.

  Similar to parsing of integers, now parse with the longest floating
  point representation "long double" via strtold (guaranteed to be
  part of C++11) and verify against the respective VGREAT values for
  overflow. Treat anything smaller than VSMALL to be zero.