[Draft] Fix free cache engine weight sync

[listar2000]

This is a tentative fix to the issue #291: where the weight sync between rollout engine and trainer engine is completely missing (an obvious issue that verl also hasn't fixed yet) when config actor_rollout_ref.rollout.free_cache_engine=False.

Fortunately this seems fixable without patching verl as in rLLM we implement our own rollout trajectory generation logic. In fact, the rollout_engine.wake_up()/sleep() method should be run regardless of the free_cache_engine flag -- the flag will be checked again anyway during the wake_up/sleep call anyway. For instance, the release of kv-cache in the inference engine will do an extra check here.

This PR should not affect the usual case when free_cache_engine = True at all. Some experiments are on-going to see if there is any side-effect when free_cache_engine = False (so this PR is still under draft)