Write an async batch processor that runs an LLM call over 50,000 documents with a concurrency limit, retries with jitter, and error isolation

You have async call_llm(doc) -> str and 50,000 documents (id, text). Write process_all(docs, call_llm, concurrency=50) that:

1. runs at most concurrency calls at once without creating 50,000 tasks in memory,
2. retries transient errors (429, 5xx, timeouts) with exponential backoff and jitter, but does not retry permanent errors,
3. isolates failures: one bad document never stops the run, and every failure is recorded with its error,
4. writes results as they finish and can resume after a crash without redoing finished documents.