Use a Symmetric matrix in the Covariant Axis2Tensor

[charleskawczynski]

This PR adds the use of symmetric matrices for the Covariant Axis2Tensors.

I think that this should be all we need, and this may result in a nice improvement in our kernels that use these tensors.

I first tried to use LinearAlgebra.Symmetric, however, it has a uplo::Char field, and this isn't a float type and as a result does not pass the check_basetype(T, S) test in our DataLayouts. So, I grabbed most of those definitions from LinearAlgebra and simply removed uplo.

I did have to widen the allowable types in AxisTensor.

Closes #1827 (the issymmetric test fails on main without fixing #1827).

[charleskawczynski]

It looks like we'll need to generalize the rmul_with_projection logic to handle more generic AxisTensors (i.e., ones backed by SimpleSymmetric matrices).

[charleskawczynski]

I'm going to leave this open, as I'm still not convinced that I did this correctly. Also, I'm not sure if and how this works on the CUDA side. This sort of optimization could account for 12 scalar fields, which is a lot of additional memory to store (in the context of columnwise!).