add deprecation warnings for old backends

[michaelosthege]

Some old backends are cluttering the codebase. We don't know of anybody still using them, so this PR adds deprecation warnings.

I'll add a line to the release notes just in case anybody ignores warnings..

[ColCarroll]

I am, in general, super in favor of this!

My only reservation is that save_trace/load_trace is currently the only non-pickle way of reloading a trace to, for example, do ppc sampling. I'm sort of ok going back to suggesting pickle unless someone has a reasonable security reason not to.

[codecov]

Codecov Report

Merging #3902 into master will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #3902      +/-   ##
==========================================
+ Coverage   83.41%   83.42%   +0.01%     
==========================================
  Files         103      103              
  Lines       14190    14203      +13     
==========================================
+ Hits        11836    11849      +13     
  Misses       2354     2354              

Impacted Files	Coverage Δ
pymc3/backends/ndarray.py	92.63% <100.00%> (+0.15%)	⬆️
pymc3/backends/sqlite.py	93.51% <100.00%> (+0.10%)	⬆️
pymc3/backends/text.py	97.16% <100.00%> (+0.11%)	⬆️
pymc3/backends/tracetab.py	100.00% <100.00%> (ø)

[michaelosthege]

I am, in general, super in favor of this!
My only reservation is that save_trace/load_trace is currently the only non-pickle way of reloading a trace to, for example, do ppc sampling. I'm sort of ok going back to suggesting pickle unless someone has a reasonable security reason not to.

sample_posterior_predictive works with InferenceData since a few weeks ago.

[ColCarroll]

😻 deprecate and merge away, then!

[rpgoldman]

If we are going to deprecate save and load trace, shouldn't we go the further step of making sample() return an InferenceData object?

It seems like we shouldn't leave things in a state where we deprecate saving and loading MultiTrace objects, but still return them.

It also seems like we should revisit @aseyboldt 's idea of making PyMC3 variable creation support a dims argument (and possibly store coords on the Model object).

[michaelosthege]

I'm all in for returning InferenceData. I still find the xarray indexing super confusing, but this is the way.

I'll do a PR tomorrow to add sample(..., return_inferencedata=False), with a FutureWarning, so we can change it to sample(..., return_inferencedata=True) in v4.0.0.
Then we'll see how much breaks (half the test suite I suppose 😅).