Fixing Python batch document read & writes

[ashvardanian]

At this point, the batch document reads and document writes are semantically wrong. They are internally replacing a batch request with a number of separate requests, which changes the expected behavior.

Worse than that, the implementation parses/converts all inputs at once, but serializes and writes them one-by-one. A proper implementation would create something like a growing_tape_t and parse+convert+serialize entries one-by-one, but would submit those for write just once!

Bringing back the ukv_format_docs_internal_k would allow us to immediately choose the optimal underlying serialization format to avoid following conversions in the backend.