feat(task): add RLinf Franka VLA pick-place deployment

[yuecideng]

Description

This PR adds the EmbodiChain side of the RLinf VLA P0 path for a Franka single-cube pick-and-place task.

It provides:

• A registered RepeatedPickPlaceRlinf-Franka-v1 task.
• A Franka VLA embodiment component with a main RGB camera.
• A 7D Cartesian action term: xyz + roll/pitch/yaw + gripper.
• Arm-only IK seeding and full-qpos reconstruction.
• Gripper normalization to the hand joint limits.
• Success detection requiring cube placement, an open gripper, and three consecutive successful control steps.
• RLinf deployment and expert Task Program configuration files.
• Task registration and package-data coverage.

The success counter now resets only for the environment rows requested by partial reset, preserving independent vectorized episodes.

Stack

• Layer: EmbodiChain environment foundation
• Base: main
• Depends on: none
• Followed by: RLinf VLA adapter PR

Issue

Refs #672

Type of change

• Bug fix
• Enhancement
• New feature
• Breaking change
• Documentation update

Validation

• EmbodiChain focused tests: 27 passed.
• Black 26.3.1 check passed for changed Python files.
• Diff and compile checks passed.
• GPU simulation and clean-process VLA inference remain environment-dependent follow-up validation.

Dependencies

The companion RLinf adapter must be installed with this task package for the end-to-end VLA smoke test.

[yuecideng]

Superseded by the merged action-contract implementation in #680 and the follow-up fix in #695.

The task and action-manager changes from this PR are already represented in the current main with the newer flat action contract. Keeping this branch open would retain the pre-#680 API and cause conflicts during rebase.