feat(memory): surface actual LLM usage/cost from inference API in sync audit

[senamakel]

Summary

The sync audit log currently estimates token usage and cost using a character-length heuristic (body.len() / 4) and hardcoded DeepSeek v4 flash pricing ($0.07/M in, $0.28/M out). The backend inference API already returns actual token counts and charged_amount_usd in the response (openhuman.usage.input_tokens, openhuman.usage.output_tokens, openhuman.billing.charged_amount_usd), but this data is lost because the memory ChatProvider trait's chat_for_text() only returns the text content.

What needs to change

1. Thread usage through the memory ChatProvider — extend ChatProvider::chat_for_text() (or add a parallel method) to return (String, Option<UsageInfo>) so callers get the actual token counts and cost.

2. Accumulate usage in summarise() — the SummaryOutput should include input_tokens, output_tokens, charged_amount_usd from the provider response.

3. Pass real usage to the audit log — ingest_summary and run_github_sync use the provider-reported values instead of estimates.

Where the data lives

• UsageInfo (with charged_amount_usd) is already parsed in src/openhuman/inference/provider/compatible.rs:707 (extract_usage)
• The ChatResponse struct at traits.rs:85 carries usage: Option<UsageInfo>
• The memory layer calls chat_for_text() which calls chat_with_history() which only returns String
• The audit log is in src/openhuman/memory_sync/sources/audit.rs

Acceptance criteria

• Sync audit entries show provider-reported token counts (not len/4)
• Sync audit entries show charged_amount_usd from the backend (not hardcoded pricing)
• Old entries (before this change) still render correctly with estimated values