Add exporter utils for model exporting (currently only implemented for PyTorch)

[Jadeiin]

Hi there! Thanks for keeping skrl so useful—really appreciate the work here. This PR introduces a lightweight PyTorch exporter utility so policies can be packaged for deployment as TorchScript or ONNX. The helper _TorchPolicyExporter wraps a policy together with optional observation and state preprocessors, exposing a simple forward(observations, states) -> actions interface for inference. Two user-facing entry points live in exporter.py: export_policy_as_jit to export TorchScript and export_policy_as_onnx to export ONNX, both handling tracing/export inputs, dummy fallbacks, device handling, and optional optimization toggles.

Some limitations:

• PyTorch backend only for these exporters
• Recurrent models (RNN/LSTM/GRU) are not supported yet
• More documentation may be needed to introduce the exporter module.

Glad to have more discussions on this PR.

[Toni-SM]

Hi @Jadeiin

Such feature is absolutely welcome. I will take a look at the implementation and get back to you with feedback.

[renanmb]

I made it work for me with limitations. The major problem was figuring out those strings. In my case I wanted to use Gaussian Mixin and the exported policy did not had as u can see below an example of what happened.

I am mentioning this because my code was fairly similar to yours so did u check if the exported policy matches ? and yes, the RNN/LSTM/GRU are quite challenging