serve: CUDA batch scheduler over a BatchEngine trait (#5279 P0) #5280

Part P0 of #5279, from the phase plan in #5279 (comment).

cuda_batch_scheduler becomes generic over a BatchEngine trait with seven methods: single request, setup+prefill, decode step, add slot, recycle slot, cleanup, reset. The dense OwnedQuantizedModelCuda implements it by plain delegation, so its behaviour does not change. This lets a Qwen3.5 engine plug into the same scheduler later (P2).

Done when:

• the scheduler builds over the trait;
• the existing scheduler tests pass unchanged;
• a new mock-engine test shows the scheduler forms Batch m=c for c in {1,4,8,16} concurrent requests.

Falsifier: the mock-engine test fails if setup ever sees m != c.