Performance regressions in linear algebra benchmarks with Hermitian and Triangular matrices

[Zentrik]

I haven't been able to replicate locally so can't bisect. But the regression is persistent on NanoSoldier so likely real.
The bad commit should be one of these ones, JuliaLang/julia@16d70b5...b8e714d.

You can find the benchmark data at https://tealquaternion.camdvr.org/compare.html?name=%28Hermitian%7CTriangular%29&regex=true&start=16d70b53f3e2493310b1bca61b3dcc861284a248&stat=min-wall-time&end=b8e714dfa3d81edd56ccda853322b6e42d94cf60 and below.
These are the min wall time results,

Summary

	Range	Mean	Count
Regressions	3.24%, 78.76%	34.89%	11

Benchmarks

Benchmark	% Change	Significance Factor
linalg.arithmetic.(exp, typename(LinearAlgebra.Hermitian), 1024)	78.76%	72.72x
linalg.arithmetic.(log, typename(LinearAlgebra.Hermitian), 1024)	78.24%	46.84x
linalg.arithmetic.(sqrt, typename(LinearAlgebra.Hermitian), 1024)	77.74%	48.91x
linalg.arithmetic.(*, typename(LinearAlgebra.UpperTriangular), Vector, 1024)	27.45%	1.21x
linalg.arithmetic.(sqrt, typename(LinearAlgebra.Hermitian), 256)	27.23%	74.58x
linalg.arithmetic.(exp, typename(LinearAlgebra.Hermitian), 256)	27.19%	66.60x
linalg.arithmetic.(log, typename(LinearAlgebra.Hermitian), 256)	27.02%	95.91x
linalg.arithmetic.(\, typename(LinearAlgebra.UpperTriangular), Vector, 1024)	25.86%	1.39x
linalg.arithmetic.(\, typename(LinearAlgebra.LowerTriangular), Vector, 256)	5.82%	1.71x
linalg.arithmetic.(\, typename(LinearAlgebra.UpperTriangular), Vector, 256)	5.22%	1.05x
linalg.arithmetic.(*, typename(LinearAlgebra.LowerTriangular), Vector, 256)	3.24%	1.52x