sample_smc conversion of sample stats to idata can result in warning

[ricardoV94]

Description

Sometimes we have less sample stats than chains with SMC (the concept of draws is hacked a bit), and this line here will trigger an arviz warning:

pymc/pymc/smc/sampling.py

Lines 266 to 270 in cc90212

	sample_stats = dict_to_dataset(
	sample_stats_dict,
	attrs=sample_settings_dict,
	library=pymc,
	)

Reported by an user in https://discourse.pymc.io/t/incorrect-number-of-draws-in-smc-example/17119/2

We should suppress that warning, (perhaps in all internal uses of dict_do_dataset by PyMC)?

CC @aloctavodia @OriolAbril

[OriolAbril]

I think it is perfectly fine to wrap those calls in ingore warnings context. I also want to float an alternative and see how we feel.

I added a check_conventions argument to the refactored dict_to_dataset. I haven't checked much how many things in PyMC will get broken by the refactor yet, but if possible it would be nice to start switching some of the internal use of arviz in PyMC to the refactored packages. Both can be installed at the same time (and will be for a bit) so this could also be a chance to get a head start on the switch (this part is basically a drop in replacement with a some extra nice things).

For now we can install arviz-base on its own, so very small python only package that depends on numpy+xarray. The only noticeable change should be a more recent xarray version as lower pin.

[ricardoV94]

Sounds good, I would like to move towards the new arviz, and possibly find a venue to make pandas / matplotlib optional, although that's a separate issue