feat: safety, quality, and robustness improvements

- bash: dangerous command blocking (9 patterns) + cd tracking across commands
- edit_file: unified diff output after edits (shows exactly what changed)
- read_file: binary file detection (rejects .png, .pyc, etc.)
- grep: skip binary files + skip .git/node_modules/__pycache__ dirs
- agent: parallel tool execution via ThreadPoolExecutor when multiple calls
- llm: fix completion_cost parameter types for streaming token tracking
- context: handle content=None in assistant messages during compression
- tests: 30 tests (was 22), added safety/diff/binary/compression tests
- README: updated line counts (~1,450), new features documented