Clear the Mamba cache table if the running queue is empty

[Wei-Lin-Intel]

This PR fixed the corner case when all the sequences in the running queue are clear, the indices of Mamba cache table would be out of bound.
Hence it fixed the accuracy issue when bs > 256. Now with bs=512, the scores of gsm8k task are stable:

vllm (pretrained=/data/Qwen3-Next-80B-A3B-Instruct,trust_remote_code=True,enable_expert_parallel=True,tensor_parallel_size=4,distributed_executor_backend=mp,max_length=16384,max_gen_toks=2048,max_num_seqs=512,max_num_prefill_seqs=16), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 512

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9386	±	0.0066
		strict-match	5	exact_match	↑	0.8878	±	0.0087

[czhu15]

LGTM