support PyStan version < 2.18

[ahartikainen]

This will first try to extract with parameters and permuted=False. In older versions, it will return numpy.array and in later versions a dictionary. If a dictionary is not seen, it will extract again, but permuted=True and then goes to reorder permutation.

Also added lp back to sample_params, it went missing after the "rebase".

[ColCarroll]

LGTM -- travis failures are:

************* Module arviz.utils.xarray_utils
arviz/utils/xarray_utils.py:310:0: C0301: Line too long (104/100) (line-too-long)
arviz/utils/xarray_utils.py:301:8: C0103: Variable name "lp" doesn't conform to snake_case naming style (invalid-name)
arviz/utils/xarray_utils.py:310:12: C0103: Variable name "lp" doesn't conform to snake_case naming style (invalid-name)
arviz/utils/xarray_utils.py:313:12: C0103: Variable name "lp" doesn't conform to snake_case naming style (invalid-name)
arviz/utils/xarray_utils.py:315:15: W0612: Unused variable 'sampler_dict' (unused-variable)

[ahartikainen]

I will still test on my local machine, after that we can merge.

[ahartikainen]

Tested with PyStan 2.18 and PyStan 2.17.1. Works now with both of them.

This means that we can start to use pip install pystan in testing.

I will add a change for that too.

[ColCarroll]

Cool!