Support variational algorithms in brms backends

[martinmodrak]

Currently not supported, but should not be that hard to pipe through to the underlying Stan backend.

[lunafazio]

I was looking into this some months ago and ended up uncovering an unfortunate interaction between cmdstanr VI output and the way that brms converts fits into rstan format regardless of backend: paul-buerkner/brms#1560

Since the error happens exclusively during summary printing of the $fit object, I don't think this should impact anything that SBC itself cares about, but just a heads up in case this comes up while developing the code.

[martinmodrak]

The current implementation of brms backends relies only on rename_pars - then it uses the inner fit$fit object to get draws. So as long as rename_pars works for the converted object, we are goood.