Stream backup archive directly to S3 to reduce ephemeral storage

[Copilot]

The daily backup job was consuming ~14 GB of node ephemeral storage because the tar.gz archive was written to local disk in full before being uploaded to S3 — making peak usage roughly raw_files + archive_size.

Changes

• aws_backup.py: Add push_stream(dest) — launches aws s3 cp - s3://... as a subprocess and returns the Popen handle for callers to stream data via stdin.
• combine_backup.py: Replace local tarfile.open(file, "x:gz") + aws.push(file) with tarfile.open(fileobj=upload_proc.stdin, mode="w:gz"), piping the archive directly to S3 as it is built. The archive is never materialized on disk.

# Before: write full archive to disk, then upload
with tarfile.open(backup_file, "x:gz") as tar:
    tar.add(...)
aws.push(backup_file, aws_file)

# After: stream archive directly to S3
upload_proc = aws.push_stream(aws_file)
with tarfile.open(fileobj=upload_proc.stdin, mode="w:gz") as tar:
    tar.add(...)

Peak ephemeral storage drops from roughly max(db_dump, backend_files) + archive_size to just max(db_dump, backend_files).

---

This change is