fix: align flashsac with g1 joystick recipe

[TATP-233]

Summary

• align the FlashSAC default recipe with the mujoco_playground G1 joystick benchmark on training-critical hyperparameters
• add a dedicated flashsac/g1_joystick/mujoco owner config and keep the collector on actor.explore(...) for FlashSAC
• switch FlashSAC reward normalization to discounted-return statistics driven by replay write order instead of sampled batch order

Testing

• UV_CACHE_DIR=.uv-cache uv run pytest tests/algos/test_flash_sac_learner.py tests/algos/test_offpolicy_worker.py tests/config/test_locomotion_params.py::test_flashsac_config_defaults tests/config/test_locomotion_params.py::test_offpolicy_flashsac_g1_joystick_task_overrides tests/scripts/test_train_scripts.py::test_offpolicy_flashsac_hydra_algo_log_name tests/scripts/test_train_scripts.py::test_offpolicy_flashsac_rejects_multi_gpu -q
• UV_CACHE_DIR=.uv-cache uv run pytest tests/config/test_locomotion_params.py::test_offpolicy_flashsac_g1_task_overrides tests/config/test_locomotion_params.py::test_offpolicy_flashsac_go1_task_overrides tests/config/test_locomotion_params.py::test_offpolicy_flashsac_g1_joystick_task_overrides tests/scripts/test_train_scripts.py::test_offpolicy_flashsac_default_task_is_go1 tests/scripts/test_train_scripts.py::test_offpolicy_flashsac_hydra_algo_log_name tests/scripts/test_train_scripts.py::test_offpolicy_flashsac_rejects_multi_gpu -q