Move to dynamically-determined polynomial orders on CPU and GPU

[samuelpmishLLNL]

As discussed here: mfem_dynamic_kernel_refactor.pdf

It's possible to change the underlying finite element kernels slightly to allow for runtime-selection of polynomial order (and quadrature rule), which can reduce the number of template parameters that appear in the physics modules, and improve compile times.

Also, the GPU shape function kernels are not being used currently and have become slightly out-of-date as a result. So, the GPU implementations should be updated to be consistent as well.

https://github.com/LLNL/serac/pull/880