Add structured test coverage and reliability improvements

[zyc-bit]

Summary

This PR introduces a structured validation stack for RLightning and improves reliability around checkpointing, weight synchronization, and environment tooling.

It adds broad automated coverage across core components and example workflows, while tightening the scripts, packaging, and supporting infrastructure needed to run those flows consistently.

Highlights

Structured Test Coverage

• Add comprehensive unit coverage across buffers, policies, engines, storage, placement, routing, and utility layers
• Add integration coverage for Ray-backed workflows, environment orchestration, remote environment paths, ALE, Mujoco, and IsaacLab-related components
• Add end-to-end smoke coverage for:
  • minimal training flows
  • OpenPI PPO
  • OpenVLA PPO
  • WBC tracking

Reliability Improvements

• Fix checkpoint saving for offloaded policies
• Improve checkpoint serialization behavior to avoid CPU aliasing issues
• Refine Async/Sync engine checkpoint and weight synchronization behavior

Tooling and Infrastructure

• Add testing scripts for full-suite execution and integration environment setup
• Add a unit-test GitHub Actions workflow
• Add Piper environment support and related env-server launch/config files
• Update dependency and packaging configuration required by the new test stack
• Add example-level uv configuration for WBC tracking and rsl_rl wiring used by WBC-related flows
• Add testing documentation for local setup and execution

Validation

Validated locally:

• .venv/bin/python -m pytest tests/unit -q
• Result: 311 passed, 2 warnings

Relevant validation entrypoints:

• bash scripts/testing/run_all_tests.sh unit
• bash scripts/testing/setup_integration_env.sh core
• bash scripts/testing/run_integration_tests.sh core -q
• bash scripts/testing/setup_integration_env.sh isaaclab
• bash scripts/testing/run_integration_tests.sh isaaclab -q
• bash scripts/testing/run_all_tests.sh e2e

Notes

• E2E coverage in this PR is smoke-only
• The E2E tests verify launch, completion, and expected checkpoint artifact generation
• They do not enforce short-run checkpoint numeric regression

[zyc-bit]

@FrankYzy @javey-q review please. And remember merge with squash.