Implement Model.debug() helper

[ricardoV94]

Example:

import pymc as pm

with pm.Model() as m:
    x = pm.Normal("x", [1, -1, 1])
    y = pm.Normal("y", mu=x, sigma=x * 2)
m.debug()

Outputs:

point={'x': array([ 1., -1.,  1.]), 'y': array([ 1., -1.,  1.])}

The variable y has the following parameters:
0: x [id A] <TensorType(float64, (3,))>
1: Elemwise{mul,no_inplace} [id B] <TensorType(float64, (3,))>
 |TensorConstant{(1,) of 2.0} [id C] <TensorType(float64, (1,))>
 |x [id A] <TensorType(float64, (3,))>
The parameters evaluate to:
0: [ 1. -1.  1.]
1: [ 2. -2.  2.]
This does not respect one of the following constraints: sigma > 0

Another example:

import pymc as pm

with pm.Model() as m:
    theta = pm.Uniform('theta', lower=0, upper=1)
    y = pm.Uniform('y', lower=0, upper=theta, observed=[0.49, 0.27, 0.53, 0.19])
m.debug()    

point={'theta_interval__': array(0.)}

The variable y has the following parameters:
0: TensorConstant{0.0} [id A] <TensorType(float64, ())>
1: Elemwise{sigmoid,no_inplace} [id B] <TensorType(float64, ())> 'theta'
 |theta_interval__ [id C] <TensorType(float64, ())>
The parameters evaluate to:
0: 0.0
1: 0.5
Some of the observed values of variable y are associated with a non-finite logp:
 value = 0.53 -> logp = -inf

Closes #4205

[codecov]

Codecov Report

Merging #6634 (d54e826) into main (3f2a1da) will increase coverage by 0.01%.
The diff coverage is 95.12%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6634      +/-   ##
==========================================
+ Coverage   91.94%   91.96%   +0.01%     
==========================================
  Files          94       94              
  Lines       15845    15923      +78     
==========================================
+ Hits        14569    14643      +74     
- Misses       1276     1280       +4     

Impacted Files	Coverage Δ
pymc/model.py	90.03% <94.87%> (+0.51%)	⬆️
pymc/distributions/continuous.py	97.70% <100.00%> (ø)

[michaelosthege]

Very cool!!

[ricardoV94]

I refined it to only show the value-evaluation pairs that lead to non-finite results.