Optimize slerp() as proposed by Gopinath Vasalamarri.

Replaces !1881 (merged)

I want to add a short note to explain why this optimization is not a significant degradation in accuracy, since this question came up in a related context.

In short, using float as an example, this MR replaces

float cos_x = ...
t = arccos(x);
sin_x = sin(t);

with

float cos_x = ...
sin_x = sqrt(1.0f - cos_x * cos_x);

One might be concerned that the latter formula is less accurate when cos_x is near 1, i.e. when cos_x = cos(theta) for a tiny angle theta near zero. However, the accuracy in the original computation is inherently limited by the representation error of cos(x) near 1, which is 2^-24 or ~5e-8. The following code snippet plots the relative error of the two formulas computed in float, compared to the original formula computed in double precision:

https://godbolt.org/z/vd1jM16s1

A subset of the table for cos(x) near 1.0:

Argument        Relative error  
cos(x)    sin(acos(x)) sqrt(1-x^2)
0.9999990  -2.703e-08   2.183e-07
0.9900980  -5.809e-08   4.806e-08
0.9802951   1.467e-08   2.410e-07
0.9705892   1.515e-09   1.253e-07
0.9609793   2.488e-08   2.488e-08
0.9514647  -1.930e-08  -1.161e-07
0.9420443  -1.479e-08   1.629e-07
0.9327171  -5.311e-08   2.954e-08
0.9234822   1.286e-08   1.286e-08
0.9143388  -3.796e-08   1.092e-07
0.9052860   2.980e-08   2.980e-08
0.8963227  -1.029e-08  -1.029e-08
0.8874483   5.239e-09   5.239e-09
0.8786616  -3.066e-09  -3.066e-09
...

As can be seen from the table, the relative accuracy of sin(t) computed from the fast formula is still accurate to a few ULPs near 1.