Refactor data read

[cdeil]

The data read should be refactored so that reading overall only opens the file once, not twice or more fore header and data read. The conversion from int to datetime should be moved to a property so that it's optional, e.g. if one just wants to write the int times exporting Whisper to Parquet. An example for this dump to Parquet should be added to the README.

whisper-pandas/whisper_pandas.py

Lines 208 to 239 in d35a1fe

	def read_whisper_archive(
	path: str, info: WhisperArchiveMeta, dtype: str = "float32"
	) -> pd.Series:
	data = np.fromfile(path, dtype=FMT_POINT, count=info.points, offset=info.offset)
	# That's right, señor. We remove all points with `time==0`.
	# The spec doesn't say, but apparamento this is what it takes.
	data = data[data["time"] != 0]
	# The type cast for the values is needed to avoid this error later on
	# ValueError: Big-endian buffer not supported on little-endian compiler
	val = data["val"].astype(dtype)
	# Workaround for a performance bug on pandas versions < 1.3
	# https://github.com/pandas-dev/pandas/issues/42606
	# int32 max value can represent times up to year 2038
	index = data["time"].astype("int32")
	index = pd.to_datetime(index, unit="s", utc=True)
	return pd.Series(val, index).sort_index()
	def read_whisper_archive_dataframe(
	path: str, archive_id: int, dtype: str = "float32"
	) -> pd.DataFrame:
	info = WhisperFileMeta.read(path).archives[archive_id]
	data = np.fromfile(path, dtype=FMT_POINT, count=info.points, offset=info.offset)
	data = data[data["time"] != 0]
	value = data["val"].astype(dtype)
	time = data["time"].astype("uint32")
	df = pd.DataFrame({"timestamp": time, "value": value}).sort_values("timestamp")
	return df

[cdeil]

Partially done in d616adf

Next the WhisperArchive as_numpy, as_dataframe, as_series should be refactored to offer flexible access to any data format the user wants (including raw e.g. for debugging, but in correct byte order)