refactor: simplify training and inference interfaces

- avoid constructing model_fn more than once when reading config
- keep inference package exports focused on public entry points
- rename extra strategy arguments to strategy_kwargs

tests/trainer/test_grpo_e2e.py

@@ -107,7 +107,7 @@ def test_online_grpo_end_to_end(base_test_env):
         device_type=device,
         nprocs=1,
         parallel_mode="none",
-        extra_kwargs={"clip_eps": 0.2, "kl_coef": 0.01, "group_size": 2},
+        strategy_kwargs={"clip_eps": 0.2, "kl_coef": 0.01, "group_size": 2},
         rollout_interval=1,
         rollout_temperature=1.0,
         rollout_top_k=0,

tests/trainer/test_online_e2e.py

@@ -109,7 +109,7 @@ def test_online_dpo_end_to_end(base_test_env):
         device_type=device,
         nprocs=1,
         parallel_mode="none",
-        extra_kwargs={"beta": 0.1, "group_size": 2},
+        strategy_kwargs={"beta": 0.1, "group_size": 2},
         rollout_interval=1,
         rollout_temperature=1.0,
         rollout_top_k=0,