Add DualModePolicyConfig, simplify inference CLI

[brentyi]

Followup to #73, which adds a dual policy inference mode that can be quickly switched between.

In #73, this was implemented by:

• Adding a secondary: InferenceConfig field to the InferenceConfig class:

  @dataclass(frozen=True)
  class InferenceConfig:
       robot: ...
       observation: ...
       task: ...
       secondary: InferenceConfig | None  # New!

• Introducing an argparse-based "pre-parser" for choosing between secondary configs, before the standard tyro parser is used to override config defaults.

Two problems here:

• (Nit) The datastructure is a bit weird, since if we have inference_config there's a inference_config.secondary, but that secondary can also have a secondary: inference_config.secondary.secondary.
• The argparse parsing logic adds complexity.

This PR:

• Adds a DualModePolicyConfig class with two fields, primary: InferenceConfig and secondary: InferenceConfig. I think this is clearer.
  • Breaking change: entrypoint-registered inference configs should now be DualModePolicyConfig instead of InferenceConfig. We could add an automatic cast / compatibility shim if this is a problem.
• Uses the DualModePolicyConfig to re-implement the secondary config selection logic using tyro.
  • This was designed to reproduce the previous logic as closely as possible. Subcommands like inference:g1-29dof-loco now refer to a (primary, secondary) inference config pair. The default secondary can be overridden with either none or another inference config.
  • The syntax is slightly different.
    • --secondary none is now secondary:none.
    • --secondary-preset g1-29dof-wbt is now secondary:g1-29dof-wbt.

@tomasz-lewicki @carlosferrazza