[BUG] Fix inference test on subset of segments for DeepARModel

[Mr-Geekman]

🐛 Bug Report

We have failing test in tests/test_modes/test_inference : test_forecast_subset_segments_failed_assertion_error.

We want to figure out why is it failing and try to fix this.
If we understand that it cannot be easily fixed, add link to this issue near the tests and explain the problem inside the issue.

Expected behavior

Tests are passed.

In this test we check that model can generate predictions on subset of segments used during training.

How To Reproduce

Run the tests.

Environment

No response

Additional context

No response

Checklist

• Bug appears at the latest library version

[Mr-Geekman]

If I try to add n_samples=None like suggested in #802 , test for DeepARModel is working without AssertionError. However, it doesn't seem to work correctly because in this mode many tests in test_models/nn/test_deepar.py fail.

Probably, it isn't a correct way to make predictions deterministic. As I understand, with n_samples=None we make model predict values in non-autoregressive fashion and take true values as input. We usually expect this during training. But during prediction it isn't really possible because true values aren't available and predictions become poor.

Setting n_samples=1 doens't fix the test.

[Mr-Geekman]

Closed by #1115.