[Roadmap] Refactor Visuomotor Control Evaluation

[siddk]

Currently, both the Franka Kitchen & Adroit Visuomotor Control Tasks are implemented on top of the R3M evaluation code. While a fairly straightforward addition to that codebase, it would be nice to:

• Programmatically download demonstration data + setup simulation dependencies (via pip)
• Refactor existing training/online evaluation to fit with the general V-Evaluation Harness API
• Add simple argparse layer for specifying splits/evaluation subsets (given 7 environments, 3 demonstration splits, etc.)

Targeting refactor by 3/4/23; happy to prioritize if folks want this sooner!