In ce9c17f3 the RV was changed such that `p` can be multi-dimensional instead of just 1-dimensional.
Since then the support length is determined by the last dimension of `p`.

This makes the outer `FunctionGraph` storage available for (re)use by inner
`FunctionGraphs` (e.g. shared variable values).

`AdvancedSubtensor` will call out to Python (i.e. Numba's object mode), because
that's not currently supported in nopython mode.  This causes problems with
indices that contain slices, because Numba has its own internal representation
of slices and no unboxing for that (i.e. conversion back to Python).

Also check exception messages in the tests.