Fix convert_to_parquet crash on empty struct fields

[Zeroupper]

Problem

convert_to_parquet() crashes when the JSON data contains empty dict fields (e.g. "metadata": {}).

PyArrow infers {} as struct<> (a struct with zero child fields), which Parquet cannot represent. This affects data types like completedapptask and image that contain empty metadata dicts.

Error log and stacktrace

Converting: 655902it [00:12, 50554.63it/s]

Traceback (most recent call last):
  File "src/carp/reader.py", line 1034, in convert_to_parquet
    self._flush_buffer_to_parquet(
        safe_name, buffers[safe_name], writers, output_dir
    )
  File "src/carp/reader.py", line 1071, in _flush_buffer_to_parquet
    writers[name] = pq.ParquetWriter(file_path, table.schema)
  File "pyarrow/parquet/core.py", line 1070, in __init__
    self.writer = _parquet.ParquetWriter(
        sink, schema, ...)
  File "pyarrow/_parquet.pyx", line 2363, in pyarrow._parquet.ParquetWriter.__cinit__
  File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
  File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowNotImplementedError: Cannot write struct type 'metadata' with no child field to Parquet. Consider adding a dummy child field.

Root cause

Some items in the JSON (e.g. completedapptask, image) contain fields like:

"taskData": { "metadata": {} }

pa.Table.from_pylist() infers this as metadata: struct<>, which pq.ParquetWriter rejects because Parquet has no representation for an empty struct.

Additionally, the previous schema handling silently swallowed cast failures (except: pass), which could lead to silent data loss when schemas drifted across batches (e.g. location type with varying fields).

Fix

• Replace pa.Table.from_pylist(buffer) with pd.json_normalize(buffer) + pa.Table.from_pandas(df). This flattens nested dicts into dot-separated columns (e.g. measurement.data.steps), avoiding empty struct inference entirely.
• Replace the silent cast-and-pass schema mismatch handling with pa.unify_schemas(), which merges schemas across batches so new columns get nulls in earlier rows instead of being silently dropped.