feat: add rate limiting, 429 retry, and error recovery for LLM calls

[prthm2910]

Summary

Adds client-side rate limiting, concurrency control, 429 detection with Retry-After header parsing, exponential backoff with jitter, and cascading KeyError prevention when LLM responses are malformed.

Changes

• SlidingWindowRateLimiter: Prevents API burst overload with configurable RPM and minimum spacing
• Concurrency semaphore: Caps concurrent async LLM calls to avoid flooding endpoints
• 429 detection: Parses provider rate-limit headers and falls back to exponential backoff
• Graceful degradation: generate_toc_init, generate_toc_continue, toc_detector_single_page, and 3 other functions now return defaults instead of crashing with KeyError
• CLI flags: --rpm-limit and --max-concurrent-requests for both PDF and markdown paths
• Config defaults: max_concurrent_requests: 0 and rpm_limit: 0 (disabled by default, opt-in)
• 40 tests covering the rate limiter, 429 handling, concurrency control, and error recovery

Closes #283

[prthm2910]

Testing summary

All 40 tests pass against the current codebase:

Test suite	Tests	What it covers
test_rate_limiter.py	8	SlidingWindowRateLimiter timing, spacing, sync/async, disabled states
test_concurrency.py	7	Semaphore limits, no deadlock, settings management, lazy init
test_429_handling.py	16	429 detection, Retry-After parsing, exponential backoff, jitter, retry exhaustion
test_error_recovery.py	9	KeyError prevention in toc_detector, generate_toc_init/continue, add_page_number_to_toc

$ pytest tests/ -v
...
40 passed in 2.39s

Also verified ConfigLoader accepts the new max_concurrent_requests and rpm_limit keys correctly (defaults to 0 = disabled, opt-in).