Unpin pytorch-forecasting dependency on specific version

[martins0n]

🚀 Feature Request

Currently we pin specific version of pytorch-forecasting (0.8.5 to be exact).
Probably we should specify only minimum and maximum versions

Motivation

Library installation in existing environments would cause less dependency problems.

Proposal

• Check new releases of pytorch-forecasting for api updates and bug fixes
• If there's no radical changes it seems we can unpeen version

Test cases

No response

Alternatives

No response

Additional context

No response

Checklist

• I discussed this issue with ETNA Team

[alex-hse-repository]

1. The next version after 0.8.5 is 0.9.0. which changes the API see. Both changes influence PytorchForecastingTransform constructor. So, as I understand, we can only use 0.8.5 or ^0.9.0 with changing the PytorchForecastingTransform constructor
2. In addition since 0.9.1 python ^3.7.1 is required, when we have ">=3.7.0, <3.10.0"
3. Also, since 0.9.1 pandas ^1.3.0 is required

Create a draft with changes for ^0.9.0 versions