Adding mixin classes for pyro training and posterior sampling

[vitkl]

This PR adds mixin classes for

1. For pyro training
2. posterior sampling.

It also fixes an issue where pyro model history was erased on load

Fixes #1031

[codecov]

Codecov Report

Attention: Patch coverage is 88.46154% with 15 lines in your changes missing coverage. Please review.

Project coverage is 90.64%. Comparing base (a0a6089) to head (52759bb).

Files	Patch %	Lines
scvi/model/base/_pyromixin.py	92.79%	8 Missing ⚠️
scvi/module/base/_base_module.py	28.57%	5 Missing ⚠️
scvi/model/base/_base_model.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1059      +/-   ##
==========================================
- Coverage   90.67%   90.64%   -0.03%     
==========================================
  Files          90       91       +1     
  Lines        6745     6885     +140     
==========================================
+ Hits         6116     6241     +125     
- Misses        629      644      +15     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[vitkl]

Interestingly, the last test works but cell2location with full data and non-amortised inference gives an error which I do not quite understand. See HTML attached (can be reproduced by taking this notebook to colab).
cell2location_analysis_3_scvi_ln_full_data.html.zip

UPD: Problem solved (a48d8ff).

[vitkl]

Sorry for the accidental merge again, I don't get why this merge happened. I am trying to merge pyro-mixin into pyro-cell2location branch using GitHub pull request web interface. For some reason, that seems to create a merge commit from pyro-cell2location into pyro-mixin.

UPD: fixed now

[vitkl]

@adamgayoso after talking to @martinjankowiak I think the default optimiser should be Adam, not ClippedAdam. I can change that. Should I modify this branch/PR or create a new one?