init_nuts not used when CompoundStep is used

[ricardoV94]

Work in #5622 revealed this issue that we seem to have had for a long time:

import pymc as pm

with pm.Model() as m:
  x = pm.Normal("x")
  y = pm.Bernoulli("y", p=0.5)
  trace = pm.sample()   # <- will not use `init_nuts` to initialize NUTS

This is actually enforced in here:

pymc/pymc/sampling.py

Lines 2278 to 2279 in 80b5e86

	if set(vars) != set(model.value_vars):
	raise ValueError("Must use init_nuts on all variables of a model.")

It seems that we should find a solution to initialize nuts robustly when it only applies to a subset of variables, unless there are good reasons not to.

[ricardoV94]

@aseyboldt tells me this ain't so bad, it simply defaults to QuadPotentialDiagAdapt.

We could allow init_nuts to distinguish between methods that can cope with subsets of variables, but that's probably low priority work