Support Translator with persistent cpu cache

[anterart]

This change allows to create instance with persistent cpu cache improving performance of Translator.unload_model(to_cpu=True) method.
With this code change, in order to create such an instance one would need to call the constructor with persist_cpu_cache=True arg, the default value of this arg is False and it keeps the same behavior as before.
This is achieved by creating a clone of the models and placing them in CPU when the constructor or the Translator.load_model methods are called.
Note that when creating an instance of translator with persist_cpu_cache=True, there will be an additional overhead for creating the models clone and placing it in CPU memory.

This change optimizes the execution of the Translator.unload_model(to_cpu=True) model and is useful in use cases where the Translator is often unloaded and loaded.

Currently the execution of Translator.unload_model(to_cpu=True) of flant5 model takes ~5.5 seconds. With this improvement it takes just about ~0.3 seconds.

[anterart]

I updated the PR, now instead of adding an additional argument to the Translator constructor, I added an optional argument keep_cache with default value false to the method Translator.load_model.
Currently there are no tests for this additional functionality, but I did test it manually.
In case the change is welcomed, I can add tests.

[NeonBohdan]

Can this feature be supported for decoder only models too?
Models swapping is very cool

Or them are too big and any way too long to load