Check to see if the log_likelihood groups have more than one data variable, and if so, require the var_name parameter. This avoids having a less understandable error message crop up later.
Introduce `var_name` parameter, and pass it through to the IC function invoked by compare.

Previously, if we got a TypeError when trying to find a log_likelihood
in from_pymc3() that TypeError would be squashed completely. Now we
will echo it to the log before handling it.

Took @OriolAbril correction.

Catch errors from IC functions invoked inside compare and annotate
them with information about the source `InferenceData` object.

Many examples show this as a good variable name for exceptions,
particularly in "except <ExceptionClass> as e:"

Error now caught sooner.

Make sure we check for multiple observed variables in compare() and that support for "var_name" works.

Previously, we checked for `sample_stats` in `get_log_likelihood()` *before* reading `log_likelihood`.  Add a check that `log_likelihood` must be missing before we check `sample_stats`.

* Limit recomputation in tests by scoping a fixture.
* Test for expected IC in `compare` test.
* Refine type assertion.