Supports passing in more MTP method types, when speculative decoding is enabled.

[changsci]

❓ Is your enhancement related to a problem?

For speculative decoding, when the algorithm type is set to MTP, it defaults to deepseek_mtp. However, the variety of models supporting MTP is currently expanding.

        vllm_speculative_algorithm_mapping = {
            SpeculativeAlgorithmEnum.EAGLE3: "eagle3",
            SpeculativeAlgorithmEnum.MTP: "deepseek_mtp",
            SpeculativeAlgorithmEnum.NGRAM: "ngram",
        }

For example, the newly released Qwen3.5 uses qwen3_next_mtp as its MTP type:

--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

💡 Describe the solution you'd like

No response

[changsci]

method xxx_mtp is deprecated and replaced with mtp in vllm v0.16.0 (see vllm/vllm/config/speculative.py).
For example, when the method is set to 'mtp', it will be automatically populated with 'qwen3_next_mtp' based on the model architecture.

[shaguaerdai]

verified in dev 8552ac1