disRNN model recovery

A place to keep Han's roadmap for disRNN.

### disRNN on nonstationary behavior
Animal behavior is nonstationary, which limits the reliability of session-wise cross-validation when fitting disRNN models. I plan to run in silico experiments to study the effects of nonstationarity on disRNN performance and explore potential mitigation strategies.

- [x] Across-session nonstationarity
- [x] Quantify its impact on disRNN fitting
- [x] https://github.com/AllenNeuralDynamics/aind-disrnn-wrapper/issues/7
- [x] Introduce session-wise variability in model params (bias, learning rate, softmax temp, etc.) and assess disRNN performance https://github.com/AllenNeuralDynamics/aind-disrnn-pipeline/issues/10
- [x] Add baselines
- [x] the ground truth likelihood per trial https://github.com/AllenNeuralDynamics/aind-disrnn-wrapper/issues/30
- [x] the RL fitted LPT https://github.com/AllenNeuralDynamics/aind-disrnn-pipeline/issues/12
- [x] Session embedding #13
- [x] Subject embedding #14
- [ ] Within-session nonstationarity #15
- [ ] Discrete states within a session
- [ ] Drifting params within a session

### Shared infrastructure
- [x] https://github.com/AllenNeuralDynamics/aind-disrnn-wrapper/pull/15
- [ ] Visualize bottleneck sigma and m over steps
- [ ] ~~Early stopping mechanism?~~