Dynamically set number of batches to appropriate value

[tyarkoni]

Per #1606, this just sets the number of batches used in statfuncs to the number of samples if smaller than 100. I made the default value None now, as it seemed like bad practice to silently replace an int value with a smaller one, but if you'd prefer to make the default 100, and raise a warning if a smaller value is used, I can tweak this.

[ColCarroll]

I think trace is a dictionary here -- could this happen down after flat_vals is defined?

[tyarkoni]

We could do that, but then batches could potentially take on a different value for each variable, which seems problematic (and would also require defining the stat func lambdas inside the varnames loop, since the values are bound by then). I don't really see a problem with doing it this way, since this is really just designed to catch the corner case where someone is testing or debugging a model with a tiny number of samples. It's a safe approach, in the sense that len(trace) is guaranteed to give a smaller value than the number of flattened samples. So I'd prefer to keep it this way...

[ColCarroll]

ah, sorry, I had a bad memory -- I knew trace was a dictionary-like object, so I thought len would be the number of keys. Turns out it is overridden, so you're getting what is intended.

[ColCarroll]

Thanks for this by the way -- I'd run into the same problem with models sampling slowly, then failing to print a summary when I cancelled them after fewer than 100 samples.