Emit canonical Claude cache metrics

[Stephen Belanger (Qard)]

Summary

• emit Braintrust canonical cache metrics from Claude Code usage instead of raw Anthropic cache metric names
• make Anthropic prompt and total token counts inclusive of cache read/write tokens
• preserve aggregate-only fallback and prefer 5m/1h split cache write metrics when Claude Code exposes the TTL breakdown
• document the canonical metric shape and update fixture/unit assertions

Validation

• bash plugins/trace-claude-code/test/run_tests.sh test_common test_post_tool_use test_fixture_replay test_stop_hook
• make test