feat: add --debug and --debug-log flags for diagnostic logging

[udid-aws]

Problem

When debugging why skill-eval functional or skill-eval trigger produces unexpected results, there's no way to see what's happening internally — which Claude command was invoked, what workspace was used, what Claude returned, or how trigger signals were classified.

Solution

Adds two global CLI flags:

• --debug — emits DEBUG logs to stderr (visible in terminal)
• --debug-log FILE — writes DEBUG logs to a file (terminal stays quiet)
• Both can be combined

What gets logged

• agent_runner: Full Claude command, working directory, exit code, elapsed time, stdout/stderr
• functional: Skill path, eval parameters, workspace paths, file copy operations
• trigger: Eval parameters, each query, trigger signal classification, tool calls, text output preview

Design

• Zero external dependencies (Python stdlib logging)
• Logs go to stderr, not stdout — tool output (reports, JSON) stays pipeable
• File uses overwrite mode (mode="w") — each run starts fresh
• Without either flag, behavior is identical to before — no logs emitted, no files created, no output difference
• Audit's existing --verbose / -v flag is completely unaffected (it controls finding visibility in audit reports, not diagnostic logging)

Usage

skill-eval --debug functional myskill              # DEBUG to terminal
skill-eval --debug-log debug.log trigger myskill   # DEBUG to file only
skill-eval --debug --debug-log out.log report .    # both
skill-eval audit myskill -v                        # audit verbose unchanged

Testing

• All 645 tests pass (8 pre-existing failures in test_config.py unrelated to this change)
• Manually verified --debug and --debug-log with skill-eval functional and skill-eval trigger against real skills
• Verified skill-eval audit --verbose behavior is unchanged with and without the new flags

Fixes #9