Speed up train/test splits for `EvaluationWindow`

[shchur]

Currently we use Dataset.map to select the past/future data for each evaluation window

fev/src/fev/task.py

Lines 191 to 213 in b091e06

	past_data = dataset.map(
	_select_past,
	fn_kwargs=dict(
	columns_to_slice=columns_to_slice,
	timestamp_column=self.timestamp_column,
	cutoff=self.cutoff,
	max_context_length=self.max_context_length,
	),
	num_proc=min(num_proc, len(dataset)),
	desc="Selecting past data",
	)
	future_data = dataset.map(
	_select_future,
	fn_kwargs=dict(
	columns_to_slice=columns_to_slice,
	timestamp_column=self.timestamp_column,
	cutoff=self.cutoff,
	horizon=self.horizon,
	),
	num_proc=min(num_proc, len(dataset)),
	desc="Selecting future data",
	)

This caches the intermediate results on disk and increases the total runtime of the benchmark due to the inefficiency.

We could instead directly work with the ListArray https://arrow.apache.org/docs/python/generated/pyarrow.ListArray.html representing each list-like column, generating the splits more efficiently in memory