Make `eigvals` work with non-Hermitian matrices

[benzwick]

Should we be more specific by throwing a DomainError here:

StaticArrays.jl/src/eigen.jl

Line 12 in 95f2578

error("Only hermitian matrices are diagonalizable by *StaticArrays*. Non-Hermitian matrices should be converted to `Array` first.")

to make it easier to catch these types of errors?

Compare with eigen.jl in LinearAlgebra:

https://github.com/JuliaLang/julia/blob/3935491dc5028f23587ca8df6d4e0de75cf50e53/stdlib/LinearAlgebra/src/eigen.jl#L376

https://github.com/JuliaLang/julia/blob/3935491dc5028f23587ca8df6d4e0de75cf50e53/stdlib/LinearAlgebra/src/eigen.jl#L418

[c42f]

Even better, I suggest wrapping LinearAlgebra.eigvals for non-hermitian matrices to return a static array of the appropriate length and type so this is no longer an error.

[benzwick]

That could also be a good idea but in my application the matrices should always be Hermitian, in which case it is better to exit with a specific error than to compute the eigenvalues using a slower method. I could first check that the matrices are Hermitian but then this check would be performed twice (once by me and once by the call to eigvals) which will slow things down slightly (I need to compute eigenvalues of 3x3 matrices millions of times so this could add up). I guess the other option would be to use a symmetric static matrix type.