Large performance counter frequency values would cause overflows, even
when 64-bit integers were used to do the multiplication with NANOSEC.
Fix this by using floating point math instead.

This just applies part of this commit:
joyent/libuv@44ecaa7