fix(envs): use RoboCasa task horizons

[anasm266]

Summary / Motivation

RoboCasa registers different rollout horizons for each task, but LeRobot currently assigns every task a default episode length of 1000. This PR uses each task's registered horizon when --env.episode_length is omitted, while preserving an explicit episode length as a global override.

Related issues

• Closes RoboCasa evaluation uses fixed episode length instead of task-specific horizons #4034

What changed

• Changed the default RoboCasaEnv.episode_length to None.
• Lazily resolve task horizons using RoboCasa's get_task_horizon.
• Set each wrapper's _max_episode_steps from the task registry.
• Preserve explicit episode-length overrides for all selected tasks.
• Raise an actionable error when a task has no registered horizon.
• Document the default and override behavior.
• Add regression tests for multi-task horizons and explicit overrides.

Explicit --env.episode_length usage remains compatible. Only the default behavior changes.

How was this tested (or how to run locally)

Validated with Python 3.12 in a local Ubuntu Docker environment:

• tests/envs/test_robocasa.py: 3 passed
• Complete tests/envs suite: 40 passed, 11 skipped
• Complete base test suite: 1086 passed, 343 skipped, 5 warnings
• pre-commit run --all-files: passed
• git diff --check: passed

A real RoboCasa GPU rollout was not run locally. The repository's RoboCasa benchmark workflow is triggered by changes under src/lerobot/envs/**.

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green
• Community Review: I have reviewed another contributor's open PR and linked it here: fix(datasets): streaming video timestamps must be file-relative, not global #4027

Reviewer notes

The RoboCasa import remains lazy so the optional dependency is not required at module import time. The evaluator is unchanged because it already respects each wrapper's _max_episode_steps.

• AI was used to implement the code based on my design. Has been thoroughly reviewed by me and understood.

[Copilot]

Pull request overview

Uses RoboCasa’s task-specific rollout horizons by default while retaining explicit global overrides.

Changes:

• Resolves registered horizons lazily per task.
• Updates configuration defaults and documentation.
• Adds regression tests for per-task defaults and overrides.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
src/lerobot/envs/robocasa.py	Resolves and applies task horizons.
src/lerobot/envs/configs.py	Makes episode length optional by default.
tests/envs/test_robocasa.py	Tests defaults, multi-task horizons, and overrides.
docs/source/robocasa.mdx	Documents default and override behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.