Need dedicated 3x3 and 4x4 matrix inversion algorithms

[FrancoisChabot]

Everything up to 4D must be screaming fast.

[FrancoisChabot]

Do we? Do we actually need this?

The 3X3 determinant code can be shown to generate identical assembly to the general one:
https://gcc.godbolt.org/z/fIxrJs

So I think we need to validate that we need to do better before putting in the specialization

[FrancoisChabot]

Anecdotally, it would appear that glm's 4x4 algorithm does lead to smaller code (assuming I didn't fail the transcription).

https://gcc.godbolt.org/z/V1fJCM

However, Vecpp's current inversion algorithm is REALLY stupid, so it might be possible to do much better in the generic code.