Add LLM request timeout and retry with exponential backoff

[Arijit429]

Closes #228
Closes #277
Closes #13

Summary

Adds timeout, retry with exponential backoff, and structured logging to
the LLM extraction pipeline in src/llm.py.

Problem

The current requests.post() call to Ollama has:

1. No timeout — hangs forever if Ollama is slow or unresponsive,
   blocking the server thread indefinitely
2. No retry — a single transient failure (network hiccup, Ollama
   momentarily overloaded) kills the entire form fill permanently
3. No structured logging — uses print() with no log levels or
   field context, making debugging impossible in production

Changes

src/llm.py

• Added 120-second timeout to all Ollama requests
• Added retry logic (3 attempts) with exponential backoff (2s → 4s → 8s)
• Smart retry: retries timeouts/connection errors/5xx, does NOT retry 4xx
• Extracted _call_ollama() method for clean separation and testability
• Replaced all print() with logging.getLogger("fireform.llm")
• Added per-field logging so extraction failures are traceable

Testing

• Existing test suite passes without modification
• Verified health check and error handling endpoints still work
• Timeout triggers cleanly when Ollama is unreachable

Changes Summary

Change	Why
120s request timeout	Prevents infinite hangs when Ollama is unresponsive
3x retry with exponential backoff	Handles transient LLM failures gracefully
Smart retry (skip 4xx)	Client errors are permanent, retrying wastes time
_call_ollama() method	Clean, testable, reusable LLM call logic
Structured logging	Replaces print(), adds field names and log levels

Real-world impact

Ensures firefighter extraction requests never hang indefinitely when Ollama is
slow on the local machine — critical for reliable field use.