Add ability to use state+rgb+segmentation obs_mode

[AlexandreBrown]

Feature Description

Hi,
As discussed in the discord channel, it would be great to have the ability to use both state (privileged state information) and visual information (eg: rgb+segmentation).

Current Limitations

One can already use state + manually do env.render() to get rgb but it is currently not possible to get the segmentation data.

Use Case

The use case behind this is to train an expert policy using privileged state information so that we can then generate expert trajectories with rgb+segmentation data to obtain rgb segments of frames generated by an expert.

Current Workaround

The current workaround seems to be to :

1. Train a policy using state obs_mode.
2. Use the trained expert policy to collect and save trajectories using the RecordEpisode wrapper.
3. Convert the state observations to rgb+segmentation using Maniskill conversion tool.
   Ex:

python -m mani_skill.trajectory.replay_trajectory \
  --traj-path my_trajectory.h5 \
  --save-traj \
  --target-control-mode my_control_mode \
  --obs-mode rgb+segmentation \
  --use-env-states \
  --num-procs 10