build: source dependencies from conda

[tristanls]

In our current setup, we use two package managers to satisfy Monty dependencies: conda and pip. This was fine when we were not distributing Monty. However, to publish and distribute Monty, we need to pick a single source for our dependencies. Today, this needs to be conda as not all dependencies are available on pip (e.g., habitat_sim).

By using conda to source all dependencies, when we publish thousandbrainsproject::tbp.monty, we can specify dependencies that should be installed when someone installs thousandbrainsproject::tbp.monty. If we do not do this, there is no standard install method for Monty on either conda or pip. (Without a standard install method, we would be asking users to execute a custom sequence of steps; not impossible, but not as accessible as conda env create).

With the changes in this pull request, the following conda environment.yml is expected to work out of the box (once this version is published) and install all dependencies needed for Monty to run:

name: monty_lab
channels:
  - aihabitat
  - pytorch
  - pyg
  - defaults
  - conda-forge
  - thousandbrainsproject

dependencies:
  - thousandbrainsproject::tbp.monty

Note

In environment.yml above, we need to specify all the channels that tbp.monty depends on because conda dependency file does not support specifying channels (see: conda/conda-build#532). For example, even though tbp.monty depends on aihabitat::habitat-sim, in the conda dependency file this will be truncated to habitat-sim. Without specifying the necessary channels, installation will fail as habitat-sim is not in the defaults. When we specify aihabitat in the channels: property, conda install will include aihabitat::habitat-sim in the search for habitat-sim.

[tristanls]

FYI: @ramyamounir, @nielsleadholm, @hlee9212 , @scottcanoe

[codeallthethingz]

LGTM

[vkakerbeck]

Looks good :)