feat: process multiple partitions in parallel

[AzulGarza]

currently, we have the num_partitions parameter for heavy requests that allows users to partition their data into chunks of time series. this process, however, is being done sequentially, which can ve slow in some cases. a good option might be to parallelize these multiple requests. to make sequential partitioned requests we used the following decorator:

nixtla/nixtlats/nixtla_client.py

Lines 748 to 773 in 7a3e52a

	def partition_by_uid(func):
	def wrapper(self, num_partitions, **kwargs):
	if num_partitions is None or num_partitions == 1:
	return func(self, **kwargs, num_partitions=1)
	df = kwargs.pop("df")
	X_df = kwargs.pop("X_df", None)
	id_col = kwargs["id_col"]
	uids = df["unique_id"].unique()
	results_df = []
	for uids_split in np.array_split(uids, num_partitions):
	df_uids = df.query("unique_id in @uids_split")
	if X_df is not None:
	X_df_uids = X_df.query("unique_id in @uids_split")
	else:
	X_df_uids = None
	df_uids = remove_unused_categories(df_uids, col=id_col)
	X_df_uids = remove_unused_categories(X_df_uids, col=id_col)
	kwargs_uids = {"df": df_uids, **kwargs}
	if X_df_uids is not None:
	kwargs_uids["X_df"] = X_df_uids
	results_uids = func(self, **kwargs_uids, num_partitions=1)
	results_df.append(results_uids)
	results_df = pd.concat(results_df).reset_index(drop=True)
	return results_df
	return wrapper

instead of calling the methods using a for loop, this process can be optimized to call them in parallel using concurrent.future.ThreadPoolExecutor (since the process is I/O and CPU bound) or multiprocess.Pool (the election of the method should be based on empirical performance).

the election of the number of parallel processes should be elected using min(num_partitions, os.cpu_count()); this behavior must be documented in the corresponding docstrings. ideally, the pr should include before and after processing time.