Replace assert checks in logprob.scan with real error handling

[jessegrabowski]

Description

There are two annoying assert checks that I commonly run up against when writing time series models here and here. Other users are hitting them too, see for example here.

It would be much nicer to have a real error message that explains what has gone wrong and why, or at least gives some information about which parts of the graph are going wrong.

I'm not 100% sure how one ends up hitting these asserts, so I tagged this with needs info. It seems to have to do with having multiple outputs from the scan inner function in a custom dist.

[ricardoV94]

In general we should address #6691 to tell users why logprob inference failed. The assert failing is a bug though, when we can't derive the logprob we should not even reach that line, and fallback to the default uninformative RuntimeError: Logprob could not be derived