fix(title): honor reasoning_effort so thinking can be disabled on title generation

[Jeffgithub0029]

Fixes the docstring/code gap in #83390: the title call's docstring claimed thinking was disabled, but the call set neither reasoning_config nor reasoning_effort, so DeepSeek-family models ran thinking at max_tokens=64 and returned an empty content.

What:

• agent/title_generator.py: read auxiliary.title_generation.reasoning_effort and pass a provider-agnostic reasoning_config. parse_reasoning_effort("none") → {"enabled": False}, which the DeepSeek / OpenCode Go provider profiles translate into thinking.type: disabled. When unset, the provider default stands (thinking ON for DeepSeek V4) — never force-disables for providers that don't support the field.
• Tests for both the enabled/disabled and unset paths.

Note: this branch is a clean rebase onto current main; it supersedes the previously closed branch based on stale upstream.

Tests: tests/agent/test_title_generator.py — 38 passed locally.

[Jeffgithub0029]

Closing as duplicate. #85424 addresses the same root cause (title generation for reasoning models / strict local providers) more completely — it drops response_format entirely and falls back to reasoning_content, verified end-to-end. This PR's reasoning_effort carve-out overlaps that scope, so I'm deferring to #85424 to avoid duplicate-review noise.