feat(rerank): add configurable HTTP timeout for OpenAI-compatible client

[Dicoangelo]

Problem

OpenAIRerankClient hardcodes a 30-second HTTP timeout. When using local LLM servers (e.g. llama.cpp on ROCm) that require cold-start model loading, 30s is often insufficient, causing ReadTimeout errors on the first request after inactivity.

Solution

Add a configurable timeout field to RerankConfig (default 30.0, fully backwards-compatible) and thread it through OpenAIRerankClient.__init__, from_config, and the requests.post call in rerank_batch. The timeout can now be set per-environment in ov.conf:

"rerank": {
  "api_key": "...",
  "api_base": "http://localhost:8080/v1/rerank",
  "timeout": 120
}

Subsequent requests still benefit from the already-warm model; only the cold-start first call needs the longer budget.

Why a config field (not a hardcoded higher value or retries)

• A higher hardcoded value isn't flexible across server setups.
• Retry logic masks the root cause instead of fixing it.

Tests

Added tests/unit/models/rerank/test_openai_rerank_timeout.py (7 tests): default timeout, custom timeout, config default, from_config threading (custom + default), and rerank_batch passing the configured/default timeout to requests.post. New tests plus the existing rerank suite pass (14 passed).

Closes #2732