Issue with recent `initval` changes

[ricardoV94]

I found some issues with the recent initval changes in #4729, as it doesn't respect the parents initvals.

This snippet often leads to a ValueError:

with pm.Model() as model:
    mu = pm.Normal('mu', initval=100)
    alpha = pm.HalfNormal('alpha', initval=100, transform=None)
    value = pm.NegativeBinomial('value', mu=mu, alpha=alpha)

model.initial_values
# {mu: array(100., dtype=float32),
#  alpha: array(100., dtype=float32),
#  value: array(1)}

When it doesn't fail, initval is still far from the expected ~ 100:

pm.NegativeBinomial.dist(mu=100, alpha=100).eval()
# array(98)

Originally posted by @ricardoV94 in #4734 (comment)

[ricardoV94]

Also a model where pm.Flat is used as input will now raise an error unless an initval is set for those variables:

with pm.Model() as m:
    x = pm.Flat('x')
    y = pm.Normal('y', x, 1)    

NotImplementedError: Cannot sample from flat variable

[michaelosthege]

closed by #4738