feat(robotwin): add hybrid environment profile

[Sonorous281]

Summary

Add an opt-in downloads_hybrid profile to RoboTwinEnv for agent-facing, single-environment execution while preserving the existing standard profile used by RL training and VLA evaluation.

The hybrid profile exposes native RoboTwin capabilities through the existing RLinf-owned environment instance:

• dual-arm robot state
• policy and agent observations
• native episode status
• native arm planning
• qpos14 and eef16 action execution
• idempotent mutation lookup
• explicit episode reset

Motivation

Agent runtimes need a persistent RoboTwin environment with native planning, calibrated observations, explicit success state, and at-most-once mutations. These capabilities should remain part of the same RLinf/native environment ownership chain instead of introducing a second simulator service.

Design

• profile: standard remains the default and keeps the existing vectorized constructor and rollout behavior.
• profile: downloads_hybrid requires one environment and disables auto-reset.
• The profile is fixed when the environment is constructed.
• Hybrid methods forward directly to the native VectorEnv capability surface.
• Startup validates the native compatibility, asset, planner, and mutation-scope contracts.
• Debug-state access remains disabled unless explicitly enabled.
• The standard profile rejects hybrid-only methods.

Files

• Extend RoboTwinEnv with the opt-in profile and capability forwarding.
• Add a single-environment example configuration.
• Add focused contract tests for profile isolation, forwarding, mutation identifiers, and compatibility checks.
• Document both profiles in English and Chinese.

Compatibility

The change is designed to leave the standard RoboTwin profile unchanged:

• no change to native controllers
• no change to standard action preprocessing
• no change to reset, step, chunk-step, partial-reset, or auto-reset behavior
• no change to task success predicates

Validation

• git diff --check passes.
• Focused unit tests are included for the new profile and standard-profile isolation.
• The local checkout could not execute the tests because its Python environment was missing locked dependencies and network access was unavailable; this PR remains a draft pending CI and maintainer review.