Use arctan2 instead of arcsin when converting to spherical coordinates

[wedesoft]

This pull request avoids using the _AVOID_DIVIDE_BY_ZERO term in two cases.

[brandon-rhodes]

Wow! This pull request is a surprise. I had always imagined arctan2(), with its multiple arguments, being much slower than arcsin(). And the call to hypot(), which I had thought might involve an intermediate Python call, in fact looks to be implemented in C. With the result that this patch seems to only incur about a 1% speed penalty for the routine, while eliminating the use of a somewhat awkward constant.

I'm inclined, therefore, to merge. I'll wait till tomorrow in case any edge cases occur to me, but I think this is likely to land. Thanks!

[wedesoft]

Ok great. I think atan2 is numerically more stable than arcsin especially for angles close to +90 or -90 degrees where sinus is near 1.0 or -1.0.