Add methods to explicitly free model from memory

[jkawamoto]

This PR introduces a close method to Llama and _LlamaModel, allowing users to explicitly free the model from RAM/VRAM.

The previous implementation relied on the destructor of _LlamaModel to free the model. However, in Python, the timing of destructor calls is unclear: for instance, the del statement does not guarantee immediate invocation of the destructor.

This PR offers an explicit method to release the model, which works immediately and enables the user to load another model without encountering memory issues. Here is an example:

from llama_cpp import Llama

llama = Llama(...)
...
llama.close()

Additionally, this PR allows the the Llama class class to be used as a context manager via the contextlib.closing interface. Here is an example:

from contextlib import closing
from llama_cpp import Llama

with closing(Llama(...)) as llama:
    ...
    # the model will be freed here

...

[abetlen]

@jkawamoto definitely agree here, thank you for the contribution.

A few things I noticed:

• We may need to perform some runtime checks / asserts to make sure the model hasn't been closed
• Right now you're only closing the model explicitly but the context and batch (which both also hold memory) still use the__del__

I also think we should have all the internal classess support the close / context manager approach, then we can use contextlib.ExitStack in the main Llama class to ensure they're cleaned up properly.

[jkawamoto]

@abetlen Thanks for your comment.

I implemented a context manager in three internal classes and added an ExitStack to the Llama class.

Regarding the runtime check of whether the model, context, and batch are still loaded, it appears that the three inner classes verify the low-level object isn't None before freeing the object. Thus, we can safely close these objects more than once. Would it be necessary to add any additional checks?

[abetlen]

@jkawamoto I've update the use of ExitStacks slightly, additionally I removed the explicit __enter__ / __exit__ methods. Since we now have a .close() method we can use contextlib.closing from the standard library to turn the Llama class into a context manager automatically.

from contextlib import closing
from llama_cpp import Llama

with closing(Llama(...)) as llama:
    ...
    # the model will be freed here

[jkawamoto]

@abetlen I think it would be better if the Llama class implements a context manager, as this would simplify the initialization of an instance by requiring fewer input characters.

Additionally, it is still worthwhile to retain the __del__ method in Llama to prevent memory leaks in cases where the user forgets to close the object.