`<xlocnum>`: Consider accumulating all digits when parsing floating-point numbers

The current approach for parsing floating-point numbers discards some digits when the input is lengthy, because we use a stack-allocated array for recording the text representation. The approach requires carefully handling the last accumulated digit and the discarded digits (see https://github.com/microsoft/STL/issues/3375#issuecomment-1413089553 and https://github.com/microsoft/STL/pull/3364#discussion_r1170192746), and adjustment for the exponent part (see #1582).

It might be nice to accumulate the whole representation (in a dynamically allocated buffer) and then send it to strtof/strtod, which should theoretically avoid the aforementioned complexity (as said by @StephanTLavavej in https://github.com/microsoft/STL/pull/3364#issuecomment-1440939821).

However, there're some bugs in UCRT's strtof/strtod (reported in DevCom-10293606 and DevCom-10345852). If we switch to accumulating the whole representation, then these bugs seem unavoidable. Note that the approach in #3364 gives incomplete workarounds.

On the other hand, using a dynamically allocated buffer might slightly slow down the parsing.

IMO it'd be better to consider switching to the new approach after the bugs of strtof/strtod get fixed.

Start by reviewing the current floating-point parsing implementation and the referenced discussions in #3375, #1582, and #3364. Verify the reported UCRT strtof/strtod issues before choosing an approach. Done would require resolving whether whole-representation accumulation is safe and implementing a tested solution that preserves all digits without unacceptable parsing regressions.