* alf snapshot only copies alf module

* fix examples dir

* get_mode for mixture distribution

* Add comments

* Address comments

* add RandomCrop gym wrapper

* address comments

* Unify ParallelNormalProjectionNetwork and NormalProjectionNetwork

* Add make_parallel() back for normal projection network

* Add parallel FC version of make_parallel to BetaProjectionNetwork

* Implement the MixtureProjectionNetwork

* Address comments

* Unify ParallelNormalProjectionNetwork and NormalProjectionNetwork

* Add make_parallel() back for normal projection network

* Add parallel FC version of make_parallel to BetaProjectionNetwork

* VQ VAE Algorithm

* Address comments

* Fix embedding summary

* Support replace() for TensorSpec

* Add type hints for replace()

* check unused parameters

* update

* improve warning message

* Causal BC Algorithm

* Add comments

* Address comments

* Address comments

* TCN wrapper

* Address comments

* Reduce arguments

* KLD for various DiscreteRegressionLoss

* Fix unittest

* MoNet algorithm

* add train play test

* address reviews; adding more code comments

* add illustrative graph for the unet

* Use torch.div for floor div in replay buffer

* Address comments

* BC for CARLA

* Address comments

* Address more comments

* Address further comment

* Support asynchronous unroll

Perform unroll and training concurrently. This can be useful in two
situations:
1. The environment step is expensive. This can help reducing time
spending on unroll
2. Realtime training where the interaction with environment happens
in realtime. For this, we can set TrainerConfig.unroll_step_interval
to the desired interaction period.

* Fix circular import

* Address comments

* Fix failure for build.sh

Need to avoid mess up with commandline arguments

* Address comments

Also fixed a potential RuntimeError "RuntimeError: unable to open shared memory object" when many training sessions are running.

* support detach for distributions

* kl_divergence has been fixed by pytorch

* address comments

* Add MuZero configuration for MetaDrive

* Add the training curve

* Address comments

* Use LSTM for reward prediction

* Add extra rewards to metadrive

* Address comments

* add advanced Fetch envs

* address comments and add gripper orientation into obs

* address comments

* Support injected summarize function from config for summarize_rollout

* inject_summary -> custom_summary

* add OneHotCategoricalGumbelSoftmax

* fix typo

* IQL algorithm

* Some minor updates

* Remove unused arguments for actor training; state-independent std etc

* Address comments

* add discrete VAE

* also return z_mode for discrete vae

* add ST gumbel-softmax

* update

* add scheduler option for the temperature

* address comments

* RewardShifting data transformer

* Make bias a required argument

* Fix observation_spec argument