UPSTREAM PR #17808: server: improve speed of speculative decoding

[loci-dev]

Mirrored from ggml-org/llama.cpp#17808

Fix ggml-org/llama.cpp#12968

I'm testing with llama-server --fim-qwen-7b-spec but it seems like the quality degraded significantly. Not sure if this is expected (as we no longer sample single token like before)

TODO: leave a drawing here to explain how it works

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary: PR #463

Analysis Overview:
Comparison of version f4ec65a5-3bd7-43ad-b473-ceef01e93350 against base 6ffb5ba3-7159-4bcb-bcc4-0ff094d14c42 for the llama.cpp server speculative decoding optimization.

---

Summary

This PR refactors speculative decoding to batch draft tokens with the main model inference, eliminating a separate llama_decode() call per iteration. The analysis shows no measurable performance differences between versions, with all binaries reporting 0.0% power consumption change and no function-level Response Time or Throughput Time variations. The code changes are structurally sound, moving draft generation before batch construction and adding rollback logic for token management, but the optimization benefits are not captured in the static analysis environment.