fix: only send think=false to Ollama, not all custom providers

[vominh1919]

Problem

When reasoning_effort: none is set, Hermes sends extra_body["think"] = false to every custom provider endpoint. The think parameter is Ollama-specific and is rejected with HTTP 422 by non-Ollama providers (Mistral, Fireworks, Together.ai, vLLM).

Fix

Gate think=false behind Ollama detection using the same pattern as _is_ollama_glm_backend():

• self._ollama_num_ctx is set (auto-detected Ollama)
• OR base_url contains ollama
• OR base_url contains :11434 (Ollama default port)

Fixes #11237

[alt-glitch]

Competing fix with #12914 for #11237 — both scope think=False to Ollama endpoints only. This PR uses _is_ollama_glm_backend() pattern, #12914 may differ in detection approach.

[teknium1]

Thanks for isolating the provider-specific request field. The underlying bug remains on current main, but this branch predates the provider-profile request path.

Problems

• run_agent.py:5328-5331 now only forwards to agent.chat_completion_helpers.build_api_kwargs(). Registered custom providers take the profile path, which calls CustomProfile.build_api_kwargs_extras() at agent/transports/chat_completions.py:562-594.
• The live unconditional assignment is now plugins/model-providers/custom/__init__.py:52-58; it still sends extra_body["think"] = False for every disabled custom endpoint.
• tests/plugins/model_providers/test_custom_profile.py:51-64 currently asserts that unconditional behavior and does not cover a non-Ollama custom URL.

Suggested changes

• Salvage the guard into CustomProfile.build_api_kwargs_extras(), using the base_url and ollama_num_ctx context already passed by the transport (agent/transports/chat_completions.py:565-573).
• Add positive Ollama and negative remote-custom-provider regression cases there.

Automated hermes-sweeper review.