[Feat][Eagle]refactor verify phase rebase

[pengchengneo]

In this PR, we have restructured some of the logic in Verfiy and DraftExtend to ensure that the shape of the batch remains the same at all stages (especially after Verify) throughout the Eagle Forward process, reducing the number of cases that require precompile, and fixing some bugs in the forward to improve the overall acceptance rate.

Furthermore, We are doing more fine-tuning to reduce data handling for TPU-> CPUs, and optimize kernel performance, and we will add e2e CI testing in the next PR and compare the performance without MTP.

launch server

uv run python3 -u -m sgl_jax.launch_server   --model-path /models/Qwen/Qwen3-32B   --trust-remote-code   --device=tpu  --mem-fraction-static=0.8   --max-prefill-tokens=4096   --max-running-requests=16   --log-requests   --log-requests-level=2 --log-level-http=debug   --show-time-cost --decode-log-interval=1   --enable-request-time-stats-logging   --attention-backend=fa   --dtype=bfloat16  --port 30000 --host 0.0.0.0 --disable-overlap-schedule --tp-size 4 --speculative-algorithm  EAGLE3 --speculative-draft-model-path  AngelSlim/Qwen3-32B_eagle3   --skip-server-warmup --page-size 16 --speculative-eagle-topk 1 --speculative-num-steps 3 --speculative-num-draft-tokens 4 --speculative-draft-model-revision 67caf31f9062d7ab64872e0a111d499bc16cd205

test command

uv run evalscope eval   --model Qwen/Qwen3-32B   --api-url http://127.0.0.1:30000/v1/chat/completions   --datasets gsm8k   --eval-type service    --eval-batch-size 128

Accuracy Test:

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!