Upload Source Table

Register an uploaded parquet relation and hand back a presigned PUT.

One call does both server-side halves of the upload flow: mint a presigned PUT into the platform's own bucket and persist the source_table row pointing at where the bytes will land. The caller then PUTs the parquet to upload_url.

The row is a managed upload: connector_type is file_upload and credential_id is NULL — there is no per-upload connector. Reads resolve the platform's own S3 credentials (S3ConnectorService.managed_config) off that sentinel.

The parquet footer is not validated here (unlike create_source_table) — the object does not exist yet; the schema is derived on demand once the upload lands.