Improve performance of TSDataset._check_endings

[Mr-Geekman]

🚀 Feature Request

Change TSDataset._check_endings to improve performance.

Motivation

Currently there is a not very optimal code for checking endings in TSDataset. It can be critical for transform and fit_transform method in dataset with many segments.

Proposal

Remove in TSDataset.check_endings cycle over segments. For example, use indexing like self.df.loc[max_index, pd.IndexSlice[:, "target"]].

You can test performance benefit on code:

from etna.datasets import generate_ar_df

df = generate_ar_df(periods=400, start_time='2020-01-01', n_segments=30000, freq='D')
ts = TSDataset(TSDataset.to_dataset(df), freq='D')
ts._check_endings()

In notebook you can use special tools for profiling.

Test cases

Add tests on _check_endings method.

1. Add test without segments with nans at the end.
2. Add test with one segment that ends with nan.
3. Add test with all segments that end with nan.

Alternatives

No response

Additional context

No response

Checklist

• I discussed this issue with ETNA Team