What this PR does / why we need it:

The get_online_features path uses a 3-step chain to transform raw online_read results into the protobuf response:

• _convert_rows_to_protobuf - iterates all rows to build per-feature (timestamps, statuses, values) vectors
• _populate_response_from_feature_data - iterates the vectors to add feature names and version metadata
• construct_response_feature_vector - iterates again per feature to scatter values from deduplicated row positions to final output positions

Each step allocates its own intermediate lists and re-iterates the full dataset, resulting in 3x passes over the data and unnecessary memory allocations adding measurable overhead to every online serving call.

This PR replace the 3-step chain with a single fused _populate_response_from_feature_data that does everything in one pass:

• Pre-computes timestamps once per row (not per feature)
• Allocates output arrays directly at output_len (the final response size), eliminating intermediate vectors
• Uses a feat_idx_map dict for O(1) feature name lookups instead of iterating per-feature per-row
• Scatters values to their deduplicated output positions in the same loop that reads them
• Appends FeatureVector results directly to the response proto

The function signature changes from accepting pre-processed feature_data tuples to accepting raw read_rows directly from online_read, removing the need for the intermediate _convert_rows_to_protobuf step entirely.