Refactor dataset loader to support multi-source datasets

[hany606]

Summary

Refactors AMPLoader to accept motion datasets from multiple directories via a sources parameter, instead of being limited to a single dataset_path_root.

Changes

• amp_rsl_rl/utils/motion_loader.py — AMPLoader.__init__ now takes a keyword-only sources: List[Dict] parameter. Each dict has "dataset_path" (root directory) and "motions" (name→weight mapping). Removes the old dataset_path_root/datasets parameters.
• amp_rsl_rl/runners/amp_on_policy_runner.py — Passes sources=self.dataset_cfg["sources"] to AMPLoader. Updated docstring.
• example/load_amp_dataset.py — Updated to use new API.
• benchmarking/benchmark_download_and_loader.py — Updated to use new API.

Breaking Changes

AMPLoader no longer accepts dataset_path_root or datasets as arguments. Callers must pass:

AMPLoader(
    device=...,
    simulation_dt=...,
    slow_down_factor=...,
    sources=[{"dataset_path": "/path/to/dir", "motions": {"Walk": 1.0, "Run": 0.5}}],
)

Motivation

The old API forced all motions to live under a single directory. This makes it impossible to compose motion priors from different retargeting pipelines or dataset collections without copying files around. The new sources list allows mixing motions from arbitrary paths.

[isorrentino]

Hi @hany606 can you solve the conflicts and rebase on main?

[hany606]

@isorrentino Done