fix(ollama): make native think configurable via extra_body

[nicoloboschi]

Summary

Fixes Ollama native structured-output for gpt-oss models, which require a
thinking level (think cannot be False) or fact extraction fails.

The native /api/chat path previously hardcoded think=False. Instead of a
model-name heuristic ("gpt-oss" in model), this wires the existing
extra_body config surface into the native payload, so operators can enable
thinking — and tune any future native field — without a code change.

What changed

_call_ollama_native now merges the configured extra_body
(HINDSIGHT_API_LLM_EXTRA_BODY and per-op variants) into the native payload,
respecting Ollama's two-tier body shape:

• Top-level native fields (think, keep_alive, …) pass through directly.
• An "options" sub-dict merges into Ollama's generation options
  (seed, top_p, num_ctx, …), with user values winning over computed defaults.

think still defaults to False; enable it per model via config:

HINDSIGHT_API_LLM_EXTRA_BODY='{"think": "low"}'   # gpt-oss
HINDSIGHT_API_LLM_EXTRA_BODY='{"options": {"seed": 42, "top_p": 0.9}}'

This differs from the OpenAI-compatible endpoints, where the SDK flattens
everything to top-level — documented in configuration.md.

Testing

• test_ollama_native_think.py: default think=False, extra_body top-level
  override (gpt-oss path), and options sub-dict merge (not leaked top-level).
• ./scripts/hooks/lint.sh passes.

Notes

Supersedes #3263, which hardcoded think="low" if "gpt-oss" in model — brittle
(guesses the level, no opt-out, bakes one model family into the transport layer).

Fixes #3246