Reduce Moose code size by using helper functions in kernels

[mortendahl]

We're currently generating a lot of code in the kernel macros, which is likely one of the main reasons why compilation time is slow and why the resulting binary is somewhat huge.

This issue is about trying to improve on this by moving some of the logic (especially in the dispatch kernels) into (non-parametric) helper functions.

Related: https://github.com/tf-encrypted/runtime/issues/477