[Feature request] decode attention support mtp-3

[ZelinMa557]

now hpc-ops support mtp-1 and mtp-2, but sglang usually recommend using --speculative-num-steps 3 for better throughput