[GSoC 2026] docs(chatbot): latency numbers

[berardifra]

Description

Follow-up to #65 (chatbot docs): adds the citable end-to-end latency numbers from the W10 latency
benchmark to the Fine-tuning & Prompting guide ("Choosing a model").

• chatbot_tuning.md ("Choosing a model"): replaces the qualitative latency text with measured
  warm numbers (no-tool ~5–6 s, tool-backed ~30–50 s, first token ~0.3–13 s) and the one-time ~70 s
  cold-load. States the empirical finding that a 7B model (mistral) did not emit tool calls on
  this stack (it answered with invented data) — so qwen2.5:3b is the default for tool-calling
  reliability, not only speed.

The _Available from version >= 6.7.0_ availability note is already on main (added in #65), so this
PR only adds the latency numbers.

Numbers come from a live end-to-end benchmark (real Ollama, no mocks, CPU-only): qwen2.5:3b
(3.1B, Q4_K_M) vs mistral:latest (7.2B, Q4_K_M).

Refs intelowlproject/IntelOwl#3810

Checklist

• The pull request is for the branch main
• I added documentation related to existing work and referenced it (follow-up to [GSoC 2026] docs(chatbot): user, deployment, developer & fine-tuning guides #65; Refs [GSoC 2026] docs: add chatbot latency numbers + availability note IntelOwl#3810)
  • Availability is already indicated on main (_Available from version >= 6.7.0_, added in [GSoC 2026] docs(chatbot): user, deployment, developer & fine-tuning guides #65)
• N/A — this is additive documentation, not a fix, so it is opened as a PR (not pushed directly to main)