fix(openai): propagate reasoning_tokens into TokenUsage for OpenAI-compatible providers

[r266-tech]

Follow-up to merged #2356 (feat(tokens): propagate cached + thoughts tokens through return contexts).

Problem

#2356 shipped TokenUsage.thoughts_tokens "for cost attribution and prompt-cache tuning" and threaded it through retain_batch_async → operation_validator, but only the gemini provider actually populates it. The OpenAI-compatible backend — the most-used class (OpenAI o-series / gpt-5, groq, deepseek-r1, plus the NousLLM / FireworksLLM subclasses) — never reads completion_tokens_details.reasoning_tokens and never passes thoughts_tokens when building its return values:

• call() extracts cached_tokens from usage.prompt_tokens_details but constructs TokenUsage(...) without thoughts_tokens → defaults to 0.
• call_with_tools() builds LLMToolCallResult(...) with neither cached_tokens nor thoughts_tokens → both default to 0.

So for every OpenAI-compatible reasoning model, the brand-new cost-attribution field silently reports thoughts_tokens=0, and the gap widens as more reasoning models are used.

Fix

Mirror the existing cached_tokens extraction (and the gemini wiring) with a 0-safe getattr chain:

thoughts_tokens = 0
if usage and getattr(usage, "completion_tokens_details", None):
    thoughts_tokens = getattr(usage.completion_tokens_details, "reasoning_tokens", 0) or 0

• call(): extract and pass thoughts_tokens into TokenUsage(...).
• call_with_tools(): extract both cached_tokens and thoughts_tokens and pass them into LLMToolCallResult(...).

The getattr chain is 0-safe for providers that don't report completion_tokens_details (non-reasoning models, the Ollama native path), so they keep thoughts_tokens=0. NousLLM / FireworksLLM inherit the fix via subclassing.

Scoped to the OpenAI-compatible provider. anthropic_llm.py also omits thoughts_tokens, but Anthropic folds thinking into output_tokens with no separate reasoning sub-count, so it's left as an optional follow-up rather than bloating this PR.

Tests

tests/test_token_usage_cached_thoughts.py gains provider-level regression tests (mocking the OpenAI client usage object):

• call() and call_with_tools() surface reasoning_tokens → thoughts_tokens (and cached_tokens).
• no-completion_tokens_details providers keep thoughts_tokens=0 (0-safe).

The existing model-level propagation tests are unchanged.