fix(closes OPEN-12012): upload batch inferences presigned URL bug

[gustavocidornelas]

Pull Request

Summary

Fixes upload_batch_inferences(dataset_path=...) uploading CSV bytes to an object key ending in .arrow, which made every batch fail asynchronously on the backend while the client reported success.

The presigned URL was minted before the input type was known, hardcoding a .arrow object key. The dataset_path branch then reassigned object_name to .csv — too late, and dead code besides, since the storage destination comes from the already-signed URL (for FS storage, the signed token payload), not from the multipart filename field. The backend dispatches on that .arrow extension and read the CSV through read_arrow_ipc, raising ArrowInvalid inside the RQ worker. POST /inference-pipelines/{id}/data only validates that the blob exists, so it returned 200 and the SDK logged Success! Uploaded batch inferences.

Both inputs now converge on the Arrow IPC path, so the uploaded bytes always match the key the URL was signed for.

Changes

• batch_inferences.py: convert a dataset_path CSV via pd.read_csv and write it as an Arrow IPC stream, so dataset_df and dataset_path share one upload path
• batch_inferences.py: remove the dead object_name reassignment to .csv that could never affect the upload destination

Context

OPEN-12012: Upload batch inferences presigned URL bug

Testing

• Manual testing

Monitoring

• No expected impact