feat: capture and persist per-sample cache_read tokens

[jshah-do]

Summary

• Map provider cached_tokens / input_tokens_details into ModelUsage.cacheReadTokens and accumulate across turns (harbor, tau, search merge, replayed usage).
• Persist run-level cache_read_tokens and sample-level sample_*_tokens (including sample_cache_read_tokens) in parquet so per-task cost can be computed from harness artifacts.
• Keep reading older parquet files without the new column (cache_read_tokens defaults to 0).

Test plan

• Smoke a short run and confirm parquet rows include cache_read_tokens / sample_cache_read_tokens when the provider reports cache hits
• Confirm Tokens/evidence rebuild can consume the new columns once portal export lands