Use of pymc3 _model pointer will cause future problems

[rpgoldman]

Describe the bug
In data/io_pymc3/ the PyMC3Converter finds the PyMC3 model -- which it needs to interpret the trace, through a pointer hidden in the pm.MultiTrace object.

This use will not be compatible with out-of-sample prediction, which in PyMC3 requires using a thinned trace in a different model (in general, one with different predictors and different dimensions).

See

arviz/arviz/data/io_pymc3.py

Lines 45 to 59 in 3aba6c9

	# This next line is brittle and may not work forever, but is a secret
	# way to access the model from the trace.
	if trace is not None:
	self.model = self.trace._straces[0].model # pylint: disable=protected-access
	self.nchains = trace.nchains if hasattr(trace, "nchains") else 1
	self.ndraws = len(trace)
	else:
	self.model = None
	self.nchains = self.ndraws = 0
	# this permits us to get the model from command-line argument or from with model:
	try:
	self.model = self.pymc3.modelcontext(model or self.model)
	except TypeError:
	self.model = None

We have already added an explicit model argument to the translator. The next step would be to deprecate using the existing back door and move towards making the PyMC3 Model object a mandatory argument.