Fixes a precision bug was introduced in 784e6d7, which improved the
precision of lerpDouble when the extrema differed significantly in
magnitude.
_lerpInt doesn't have this issue since the extrema are passed as 64-bit
twos-complement values, therefore the difference will always be accurate
so long as it doesn't overflow. This reverts the _lerpInt implementation
to the original a + (b - a) * t, but adds a test to avoid a regression
if anyone is tempted to make it consistent with the others.
6e0262d37a