[WIP] Time series modules

[hesenp]

Some time series wrapper as discussed in #1361 . Could appreciate any discussions and inputs as the effort is very nascent.

A few things this model tries to accomplish:

1. provide a building blocks for time series analysis. i think this should not be limited to forecasting only. it should have interpretation capability as well.
2. the building blocks should fit into the fit, predict pattern like sci-kit learn.
3. hide away sampling procedures.
4. hide away some of the numpyro handler functions, which could be a bit daunting to new users (including myself)
5. still enable flexibility to try out different patterns.

With the above objectives in mind, this PR (so far) includes an implementation of a naive AR(1) model that is built on the following procedures:

• method 1: low level implementation of model, based on NUTModelTrainer only.
• method 2: higher level implementation of model, based on StateSpaceModel.

Could appreciate any inputs and hole poking on whether these patterns in method 1 and 2 would break in certain situations. I might come back and add more text book examples later.

[fehiepsi]

Maybe it is fine to not support SVI in time-series module. We can have a separate forecast module that follows Pyro semantic: Forecaster and HMCForecaster.

[hesenp]

ah sorry i'm very slow these days. will get back to it ha.