fix double rounding of subnormal hex-floats in StringToDouble

[Ramya-9353]

Repro: with ALLOW_HEX_FLOATS, StringToDouble("0xcc5f893a94ec6.a8ap-1074") returns 1.7763493645383205e-308, one ulp below the value strtod and the exact value round to (1.776349364538321e-308). Hex-floats whose significand needs more than 53 bits and whose magnitude is subnormal can be a ulp out.
Cause: RadixStringToIeee rounds the significand to 53 bits (round-to-nearest) and then forms the result with ldexp. For a subnormal result ldexp rounds a second time onto the narrower 2^-1074 subnormal grid, so the value is rounded twice. #304 fixed the same double rounding on the StringToFloat path, but the StringToDouble path still rounds to nearest at 53 bits before the subnormal ldexp.
Fix: keep the exact significand, its dropped-bit count and a sticky bit; when the result is subnormal, round the exact significand once, straight onto the 2^-1074 grid, so ldexp only has to place it. Normal results, and the float, decimal, hex-integer and octal paths, are unchanged. Checked against strtod over a wide sweep of subnormal hex-floats.