remove weight_bridge: vLLM handles TP slicing internally
vLLM 0.22.0's RLHF weight transfer API (/update_weights) accepts full
tensors and handles TP slicing on the server side. The weight_bridge
module that manually sliced tensors per TP rank is unnecessary dead code.
Remove:
- examples/opsd/opsd/weight_bridge/ (base, qwen2, qwen3)
- examples/opsd/tests/test_weight_bridge.py
Also remove unused arch= kwarg from build_rollout() call in main.py.
Signed-off-by: Guokai Ma <guokai.ma@intel.com>
