server: improve speed of speculative decoding

[ngxson]

Fix #12968

I'm testing with:

• draft model: https://huggingface.co/unsloth/Qwen3-0.6B-GGUF (using Q8_0)
• main model: https://huggingface.co/unsloth/Qwen3-8B-GGUF (using Q4_K_M)

So far the results are coherent.

How it works:

[ngxson]

server tests passed locally, this should be ready for review @ggerganov

[theo77186]

Just a cosmetic bug: in the llama-server logs, the eval time is 0.00ms, thus the total time only accounts for the prompt processing time. It also causes the eval tokens per second to be meaningless. The model outputs seem to be correct, though.

[ngxson]

@theo77186 you mean just the stdout/stderr log, right? (which is not the stats returned by API)

Edit: I think I need more details on the bug, as well as step-by-step reproduction. Feel free to open a dedicated issue.

[ggerganov]

@theo77186 you mean just the stdout/stderr log, right? (which is not the stats returned by API)

Both of them (logs and UI) were broken in cases when the draft batch was always accepted (e.g. "count from 1 to 100"). I fixed it with f74d1ee.