fix: tolerate malformed token usage values

[royalpinto007]

Fixes #5

Summary

Makes token usage parsing defensive so malformed usage values from Claude Code transcripts default to 0 instead of raising and breaking trace emission.

Details

• Adds a small _usage_int(...) helper for token usage coercion.
• Applies it to input, output, cache read, and cache creation token fields.
• Adds a regression test covering string/object malformed values while preserving the assistant text output.

Why

Transcript data is external input. A single malformed usage field should not prevent the plugin from emitting an otherwise valid trace.

Testing

• uv run pytest tests/test_tokens.py
• uv run pytest

[royalpinto007]

# Before fix
PYTHONPATH=. python /tmp/repro_traceroot_plugin_bad_usage.py

Traceback (most recent call last):
  File "/tmp/repro_traceroot_plugin_bad_usage.py", line 20, in <module>
    calls = tokens.llm_calls([msg])
            ^^^^^^^^^^^^^^^^^^^^^^^
  File "traceroot_observability/tokens.py", line 121, in llm_calls
    inp = int(u.get("input_tokens") or 0)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: invalid literal for int() with base 10: 'unknown'

# After fix
PYTHONPATH=. python /tmp/repro_traceroot_plugin_bad_usage.py

parsed successfully
prompt 0
completion 0
cache_read 0
cache_creation 0
text ok