Added warnings for non-positive support in positive parameters

[fonnesbeck]

Since so many continuous variables have positive-valued parameters, this just generates a simple warning if you pass a parameter that has non-positive support. As best it can, it checks whether the logp at zero is finite. This should help users avoid mis-parameterizing models.

[twiecki]

Good idea.

[springcoil]

Merging this.

[aloctavodia]

Now something like this

    sigma = pm.HalfNormal('sigma', sd=10)  # or pm.HalfCauchy('sigma', 10) 
    y = pm.Normal('y', mu=mu, sd=sigma, observed=data)

results in a warning message. I guess this is related to sd=0 being problematic, but given that in practice it is ok to use HalfNormal (or HalfCauchy) as a prior for sd, I think this warning is misleading.

[fonnesbeck]

I agree. I need to fine tune the warning a little.

[fonnesbeck]

@aloctavodia Try out PR #1482

[aloctavodia]

@fonnesbeck LGTM. Thanks for the prompt fix!