fix(llm): add timeout and transient failure handling to Ollama calls

[SeanClay10]

Problem

Both chat() calls in src/llm/llm_client.py (initial extraction at line 128, retry at line 187) have no timeout and no error handling around the network call itself. The current retry logic only handles null fields in a successful response — it does not handle connection errors, server errors, or hung requests.

This means:

• If the Ollama server becomes unresponsive mid-run, the pipeline hangs indefinitely with no output and no way to detect it without watching the terminal
• In multiprocessing mode (classify_extract.py --workers N), a single hung worker blocks its slot in the pool for the rest of the run
• A transient 503 or connection reset kills the entire job for that file rather than retrying

Tasks

• Wrap both chat() calls in a try/except that catches connection errors, timeouts, and malformed response errors
• Add a configurable timeout to each chat() call (e.g. --llm-timeout CLI flag, defaulting to a reasonable value like 120s)
• Implement retry with exponential backoff for transient failures (network errors, 5xx responses), separate from the existing null-field retry
• Log the raw response content on failure to aid debugging
• Ensure a hung or failed worker in multiprocessing mode produces a clear error row in the summary CSV rather than silently disappearing

Context

src/llm/llm_client.py:128 and :187 are the two call sites. The existing null-field retry at line 148 is unrelated — it re-prompts on a successful response that returned nulls, not on a failed call.