Add eval support and debug sensor

[eundersander]

Run eval on a trained model's checkpoints as follows:

python habitat_baselines/run.py --exp-config path/to/your_gala_kinematic_eval_config.yaml --run-type eval

Some notes for these changes:

off-by-one

• we should look closely at _eval_checkpoint to confirm I didn't introduce an off-by-one bug in terms of actions, steps, observations.

BatchedEnv.is_eval

• this is a hacky public member var for now but should instead be hooked up to a yaml config
• this member dictates that BatchedEnv should run through all episodes in the dataset only once; envs are disabled as necessary towards the end of the evaluation

recommended config for eval

• DEBUG_RGB_SENSOR
  • when evaluating, be sure to include "DEBUG_RGB_SENSOR" in config.SENSORS
  • be sure not to include this when training (it will hurt runtime perf)
• debug sensor size
  • see batched_env.py debug_width and debug_height
  • the size here greatly impacts eval speed
• VIDEO_OPTION: ["disk"]
  • use this for eval
  • Beware I hit a runtime error when I enabled the "tensorboard" option (maybe my version of tensorboard is outdated)
• NUM_ENVIRONMENTS: 32
  • Maybe try larger, but beware I was seeing a crash on my system seemingly due to GPU OOM.
• EVAL_CKPT_PATH_DIR
  • set to the same as CHECKPOINT_FOLDER

gala_kinematic/python/sim_test.py (not in habitat-lab)

• this is a script that generates a few random rollouts with videos (random actions; no inference)
• this is a handy place to tune cameras

[ASzot]

LGTM. I can also add in more of the rendering functionality from Habitat 2.0, like rendering all of the info metrics to the current frame so we can have better insight into metrics used to compute rewards.