On Windows, the FP stack registers are used with less precision.
This causes rounding errors in the uint64 to float32 conversion.

This CL replaces the implementation based on FP stack registers
with an implementation based on bit operations. This implementation
is 2x slower than the original implementation.

An alternative would be to change the precision of the FP stack
registers just for the uint64 to float32 conversion. However, in a
micro-benchmark this is 5-6x slower than the original implementation.
It is also not clear if changing the precision could cause side
effects.

R=clemensh@chromium.org

Change-Id: Iaab6b6f258ff01e0c6e93f3632daf516fae3e74b
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1708486
Commit-Queue: Andreas Haas <ahaas@chromium.org>
Reviewed-by: Clemens Hammacher <clemensh@chromium.org>
Cr-Commit-Position: refs/heads/master@{#62986}