Fix disk-cache LRU logic

[donbcd]

The LlamaDiskCache tried to implement LRU logic, but did not succeed. getitem did a pop() from the cache, but no push (that was commented out). This results in a miss-hit-miss-hit-... behavior if the same prompt is executed repeatedly.
setitem tried to reorder things, which does not make sense if there is always only one or zero elements.

The solution is as simple as it gets: The used "diskcache" already has implemented LRU behavior by default, so LlamaDiskCache does not need to do anything, just "get" and "set".

[Copilot]

Pull Request Overview

This PR simplifies LlamaDiskCache by removing custom LRU management and relying on the built-in LRU behavior of the underlying diskcache library.

• Replaced pop/push in __getitem__ with a simple get
• Removed manual eviction loop in __setitem__ and added close() call
• Eliminated redundant deletion logic in __setitem__

Comments suppressed due to low confidence (2)

llama_cpp/llama_cache.py:142

• [nitpick] This debug print to stderr may clutter logs in production. Consider removing it or replacing it with a proper logger call at an appropriate log level.

print("LlamaDiskCache.__setitem__: called", file=sys.stderr)

llama_cpp/llama_cache.py:144

• With the removal of manual eviction logic, add or update tests to verify that the diskcache's default LRU eviction behaves as expected under capacity pressure.

self.cache[key] = value

[Copilot]

Calling close() on the cache on every __setitem__ invocation will close the underlying DB and prevent further operations (and degrade performance). Move close() to a teardown or destructor method instead of inside the setter.

Suggested change

self.cache.close()