Optimize routed-expert kernel launches and clean up code

[MRPRESIDENT66]

Performance

Tested on an NVIDIA RTX 4090 using JDK 21, TornadoVM 5.2.1-dev, the CUDA backend, and --cuda-graphs.

Command:

./llama-tornado --gpu \
  --model Qwen1.5-MoE-A2.7B-Chat.Q8_0.gguf \
  --prompt "Explain briefly how mixture of experts routing works." \
  --max-tokens 64 \
  --temperature 0 \
  --seed 42 \
  --cuda-graphs

Five interleaved runs were performed after one warm-up run:

Version	Run 1	Run 2	Run 3	Run 4	Run 5	Mean
Before fusion	129.17	129.92	128.11	129.92	130.52	129.53 tok/s
After fusion	143.93	144.26	143.11	145.59	144.22	144.22 tok/s

The fusion improved throughput by 11.34% with CUDA Graphs enabled.

[CLAassistant]

All committers have signed the CLA.

[mikepapadim]

@MRPRESIDENT66 thank you, it LGTM. I would suggest to followup with prefill-decode batched.
Also, for a sanitiy check profile it once with JFR to analyze the host (java) for potential opts and then with nsys to check for potential gpu (cuda) opts.

You can have a look of what is supported from CUDA in tornado in beehive-lab/TornadoVM#1026.

[MRPRESIDENT66]

@mikepapadim Thanks! I’m currently working on batched prefill-decode for Qwen2-MoE. I’ll also profile it with JFR and nsys as you suggested, and check the CUDA features in #1026.