feat: add Sharpa sim2real and HORA training support

[Mingrui-Yu]

Summary

• add Sharpa in-hand PPO, APPO, HORA, and distillation configs and training entrypoints
• extend the MuJoCo backend and Sharpa env support for sim2real randomization, contact force observations, and object setup changes
• move Sharpa HORA owner/runtime resolution out of training scripts and into owner-selected runtime helpers
• expose Sharpa MuJoCo model metadata through backend contract methods instead of env-level MuJoCo internals
• keep HORA APPO shared IPC on the ADR-0005 obs/critic contract and derive HORA model priv_info from critic
• add documentation, support-matrix updates, and Sharpa-focused tests and benchmarks

Validation

• uv run pytest tests/algos/test_hora_contract.py tests/benchmark/test_sharpa_init_dr_benchmark.py tests/scripts/test_check_docs.py::test_sharpa_domain_randomization_doc_matches_owner_config -q
• uv run pytest tests/scripts/test_check_docs.py -q
• uv run pytest tests/test_sharpa.py tests/envs/test_sharpa.py tests/envs/test_env_configs.py::test_sharpa_grasp_env_initializes_dr_once_with_grasp_provider -q
• uv run pytest tests/algos/test_hora_imports.py tests/scripts/test_train_scripts.py::test_hora_distill_script_delegates_teacher_owner_resolution tests/scripts/test_train_scripts.py::test_hora_distill_task_owner_overrides_root_config_defaults tests/scripts/test_train_scripts.py::test_hora_distill_teacher_owner_defaults_support_ppo_and_appo tests/scripts/test_train_scripts.py::test_hora_distill_teacher_run_slug_omits_teacher_run_name tests/envs/test_sharpa.py::test_sharpa_env_uses_backend_contract_for_mujoco_metadata -q
• uv run pytest tests/scripts/test_train_scripts.py::test_play_wrapper_preserves_hora_priv_info_and_proprio_history tests/scripts/test_train_scripts.py::test_play_wrapper_timeout_bootstrap_preserves_hora_priv_info tests/scripts/test_train_scripts.py::test_hora_distill_task_owner_overrides_root_config_defaults tests/scripts/test_train_scripts.py::test_hora_distill_teacher_owner_defaults_support_ppo_and_appo tests/scripts/test_train_scripts.py::test_hora_distill_teacher_run_slug_omits_teacher_run_name -q
• uv run pytest tests/base/test_sim_backend_smoke.py tests/algos/test_hora_imports.py tests/scripts/test_train_scripts.py tests/test_sharpa.py tests/envs/test_sharpa.py tests/envs/test_env_configs.py tests/benchmark/test_sharpa_init_dr_benchmark.py tests/scripts/test_check_docs.py::test_sharpa_domain_randomization_doc_matches_owner_config -q
• uv run pytest tests/ipc tests/algos/test_hora_contract.py tests/algos/test_hora_imports.py tests/algos/test_appo_runner_unit.py tests/scripts/test_train_scripts.py tests/test_sharpa.py tests/envs/test_sharpa.py tests/envs/test_env_configs.py tests/benchmark/test_sharpa_init_dr_benchmark.py tests/scripts/test_check_docs.py::test_sharpa_domain_randomization_doc_matches_owner_config tests/base/test_sim_backend_smoke.py -q

Impact

• Sharpa HORA PPO runtime dispatch is now selected by owner YAML via algo.runtime_resolver; missing custom runtime resolvers fail fast.
• HORA distillation script no longer parses teacher owner YAML or owns HORA actor-class checks directly.
• Sharpa env no longer imports MuJoCo or reads self._backend.model; MuJoCo-specific metadata is accessed through SimBackend.
• APPO shared rollout storage no longer transports priv_info as an IPC field; HORA derives its internal tensor from the env-owned critic observation.
• Sharpa Motrix remains unsupported for this task because the required randomization and metadata contracts are MuJoCo-only today.

Checklist

• Owner/config logic stays outside long-lived training script rules.
• Backend-specific model metadata access is isolated behind backend contract methods.
• Docs and benchmark overrides match env.domain_rand.scale_list and Sharpa interval body_force DR.
• Targeted tests cover HORA state contract, HORA runtime selection, HORA APPO IPC, Sharpa metadata isolation, backend metadata getters, docs, and benchmark construction.