Improve support for custom integration kernels

[sclaus2]

Currently it is rather difficult to customize ffcx generated integration kernels. This issue is to outline some ideas on how to improve ffcx to make this easier. My aim is to use these improvements to write a runtime integration kernel extension to ffcx. The steps discussed so far are:

1. Introduce a 'void*' to tabulate_tensor arguments in ffcx
2. Keep details of basix elements used for tabulation in intermediate representation of integrals, i.e. prevent loss of information across ffcx.
3. Make basix tabulate order consistent between basix and ffcx generated code

First PRs connected to this are: FEniCS/dolfinx#3643 and #749 (1.) .