Lazy RTC compilation of MAGMA basis kernels for small P, Q

[sebastiangrimberg]

Refactored preliminary PR for #1301. Compiles custom non-tensor MAGMA kernels only when needed (during first call to CeedBasisApply).

Other included improvements to the magma backend:

• Renames files in the magma backend to match the other GPU backends in the library, along with other style formatting for consistency with other backends and the style guide (variable casing, for example).
• Simplify some of the macro definitions in ceed-magma.h to eliminate duplication between tensor and non-tensor basis code.
• Add include guards to jit-source/magma files as done in Style fixes for cuda and hip backend consistency #1299 for cuda and hip backends.
• Use JiT for the non-tensor weight kernels for the CEED_EVAL_WEIGHT in the magma backend (as for cuda/ref and hip/ref).

TODO

• Test if resolves Slowdown in PETSc multigrid example with tetrahedral elements and /gpu/*/magma backends #1373

[sebastiangrimberg]

Indeed this PR resolves #1373:

Before:

After:

[jeremylt]

thanks for these changes!

[sebastiangrimberg]

Looks like the only CI issue is a -0.05% coverage change (noise). So I think this can be merged when ready.