fix: derive default llm-rate from resolved --workers

[rasros]

Problem

DEFAULT_LLM_RATE = min(2 / DEFAULT_WORKERS, 0.2) was computed once at import time from os.cpu_count() and baked in as the argparse default, so it ignored the --workers the user actually chose. The design intent is "keep ~2 LLM calls in flight", but on a 20-core box with --workers 4 the rate stayed pinned at 0.10 (~0.4 concurrent calls, a much slower/cheaper run than intended); the reverse case over-shot.

Fix

• Extract _default_llm_rate(workers) = min(2/workers, 0.2).
• --llm-rate now defaults to None; after parsing, when the user did not pass it explicitly, it's derived from the resolved ns.workers. An explicit --llm-rate still wins.
• Help text updated to describe the derivation.
• DEFAULT_LLM_RATE retained (from DEFAULT_WORKERS) for --help display and back-compat.

Tests

• test_default_llm_rate_tracks_workers (0.2 cap at --workers 4; 2/40 at --workers 40).
• test_explicit_llm_rate_overrides_workers_derivation.