Fix handling of small magnitude quaternions

[qsantos]

Hello again! I contributed to the project last year in #946. I have finally updated the version of glm in my project, and I noticed that there is still actually something afoul with small-angle quaternions.

The special case handling in glm::qua::pow is specifically intended for the case when the magnitude is null. Even for magnitudes smaller than epsilon, the common formula should be used. Because comparing a floating-point value by equality triggers a warning through the -Weveverything setting of clang, it must be selectively disabled for the condition.

I initially did so in a921b51 but, as you can see, the pipeline failed because of -Wfloat-equal, and I fixed it too quickly by using a comparison to epsilon like in other parts of the project.

In passing, glm::epsilon is defined in ext/scalar_constants.inl as std::numeric_limits<T>::epsilon. According to cppreference.com, it “returns the difference between 1.0 and the next representable value of the given floating-point type”. But, for small values of x, there are many floating-point values between x and x + epsilon. Dawson's series on floating-point numbers seems to agree with me. Also, for values of x larger than 2.0, comparing to epsilon is the same as simple equality. So it is possible that other floating-point discrepancies have crept up somewhere else due to comparing to epsilon.

Tell me if this fix is acceptable (for information, I am actually using the version with if (VectorMagnitude == 0) { if (< std::numeric_limits<T>::min()) { in my project, and it works perfectly).

[qsantos]

I finally took the time of fixing the pipeline for this PR.

[Groovounet]

Thanks for contributing!