refactor(miles): drop redundant num_gpus=1 in RLix train actor decorator

[TianyeGGBond]

Context

In the F12 shared-PG / RLix-mode train path, RayTrainGroup._allocate_gpus_via_placements builds each train actor like this:

TrainRayActor = ray.remote(num_gpus=1, runtime_env={"env_vars": env_vars})(actor_impl)
actor = TrainRayActor.options(
    num_cpus=num_gpus_per_actor,
    num_gpus=num_gpus_per_actor,   # <- always overrides the decorator
    scheduling_strategy=...,
).remote(...)

.options(num_gpus=num_gpus_per_actor) is applied on every single use, so the num_gpus=1 in the ray.remote(...) decorator is never the effective value. It is dead, and worse it is misleading: a reviewer sees num_gpus=1 in the decorator and num_gpus=num_gpus_per_actor (0 in RLix mode) in .options() and has to reason about which one wins.

Change

• miles/ray/actor_group.py: remove num_gpus=1 from the ray.remote(...) decorator. GPU resource is now declared in exactly one place — .options().

Impact

None. .options() already overrode the decorator on every code path, so effective scheduling behavior is byte-for-byte identical. This is a readability-only change.