type-stable inner loop for sqrtm (#20214)
* type-stable inner loop for sqrtm
As suggested by Ralph_Smith on [discourse](https://discourse.julialang.org/t/review-schur-pade-matrix-powers-speedup/1650/6)
On my machine: speedup x15
* dispatch sqrtm on real-or-not bool
As suggested by @stevengj
* sylvester for numbers