serve: Qwen3.5 CUDA multi-slot greedy step, looped (#5279 P1) #5289

Phase P1 of #5279.

Qwen35CudaModel decodes one sequence at a time (forward_single / forward_single_greedy take one Qwen35CudaState). The serve batch scheduler (P0, #5280) needs an engine that takes one step for m sequences at once.

Change. Add Qwen35CudaModel::forward_batch_greedy(tokens, states, positions) -> Vec<u32>. It takes one greedy step for each of m slots by looping forward_single_greedy per slot. The shape is checked before any device work: the three lengths must be equal and non-zero. Later phases (P4 onward) replace the loop with batched kernels behind the same signature.

Done when.

• On the real Qwen3.5-0.8B file, for m = 1..4 with ragged prompt lengths, every slot's greedy stream from the interleaved batched steps is bitwise equal to the same prompt decoded alone with forward_single_greedy. This shows that the model's shared scratch buffers do not leak between slots.
• A shape mismatch (unequal lengths, m = 0) is refused by name before any device work.