fix(openai): stop max_tokens streaming retry loop on reasoning models (#9716)

[Dennisadira]

What

Fixes the max_tokens streaming loop reported in #9716.

When a thinking model spends its entire max_tokens budget on the reasoning block, the C++ autoparser clears the raw Response and delivers reasoning-only ChatDeltas (no content, no tool calls). ComputeChoices' empty-response retry then sees an "empty" generation and regenerates from scratch — up to maxRetries (5) times — each attempt re-consuming the whole budget. So actual token consumption becomes a multiple of max_tokens, and the stream never cleanly terminates with finish_reason: "length".

Fix

• Add a small reachedTokenBudget(completion, maxTokens) helper.
• Suppress both the built-in empty-response retry and the caller-driven (reasoning-only) retry when the completion count has reached the configured max_tokens ceiling — regenerating would just hit the same ceiling again.
• Report finish_reason: "length" (new FinishReasonLength constant) instead of "stop" in both the streaming final chunk and the non-streaming response when the budget was exhausted. Tool/function finish reasons are left untouched.

Tests

Adds a deterministic regression test that stubs ModelInferenceFunc and counts backend invocations for the budget-exhausted reasoning case:

• Before: backend invoked 6 times (1 + 5 retries)
• After: invoked 1 time

Plus boundary unit tests for reachedTokenBudget (no limit / below / at / above). Full core/http/endpoints/openai and core/backend suites pass; gofmt + go vet clean.

Scope / follow-up

This is a focused Go-side fix: budget exhaustion is derived from the completion-token count vs max_tokens, so it needs no backend/proto changes and resolves the user-facing loop today.

A natural follow-up — matching the "plumb the backend's length stop reason through" suggestion in the issue — would be to surface an authoritative stopped_limit from the llama.cpp gRPC server (and other backends) rather than inferring it. Happy to do that as a separate PR if you'd prefer the backend as the source of truth; I'd just want to confirm the proto shape (bool vs. a stop-reason enum) and default for backends that don't report it before writing it. The retry-suppression logic here is unchanged either way.

Assisted-by: Claude:claude-opus-4-8 [Claude Code]