Clean up TFP io

[junpenglao]

• Update io_tfp to TF2.
• Use official TFP API like jointdistribution* and move Edward to io_edward2
• Ordering of chain and sample (see also from_tfp with multiple chains? #456)
• Mapping of sample/diagnostics to a dict, currently it is possible with user code like https://gist.github.com/junpenglao/51cd25c6372f8d2ab3490d4af8f97401, however, the difficult here is that user can specify the diagnostic output using trace_fn, which likely does not have the same structure.
• Add tests using a multiple chains model (once from_tfp with multiple chains? #456 has been fixed)
• Make sure pointwise log likelihood is stored in log_likelihood group (if getting pointwise log likelihood is possible at all)
• Enforce ArviZ schema: group names and sample stats naming convention as much as possible

[junpenglao]

Updating the TODOs.

[OriolAbril]

I would also add to the TODOs:

• Add tests using a multiple chains model (once from_tfp with multiple chains? #456 has been fixed)

[junpenglao]

Thanks! Added.

[percygautam]

@junpenglao @OriolAbril Do we just need to upgrade TF1 to TF2 in first task? Can we follow this guide?

[OriolAbril]

Yes, this guide is great reference

[percygautam]

Okay, thanks

[percygautam]

@OriolAbril @junpenglao I have some doubts. Since there is no concept of placeholders in TF2. How are we going to work with data having placeholders? We can simply convert a tensor object to ndarray using tf.make_ndarray if it do not contain any placeholders. Kindly guide me through it.

[OriolAbril]

closing in favour of #1890