fix(openvino): explicit ov::Tensor frees in ggml_backend_openvino_free

[thedanhoffman]

Overview

The following throws a segmentation fault on exit.

export GGML_OPENVINO_DEVICE=NPU
./bin/llama-cli -m Llama-3.2-1B-Instruct-Q4_0.gguf -ngl 99 -np 1 -c 512 --simple-io --single-turn -n 1 -p hi

Root-caused to a use-after-free involving OpenVINO's ZeroMemPool. Goes like this

1. ZeroMemPool::allocate_zero_memory() returns std::shared_ptr<ZeroMem>
2. std::shared_ptr<ZeroMem> has a custom destructor method of the form [this](ZeroMem *ptr) { delete_pool_entry(ptr); } (where delete_pool_entry is a non-static member function of ZeroMemPool)
3. this is invalid by the time the destructor runs because RAII does not clean up std::shared_ptr<ZeroMem> consistently

This is downstream of std::shared_ptr<void> being used for runtime_context not propagating the destructor call to class members.

Went ahead and defined a clear_caches method that does this destruction explicitly (but not using std::shared_ptr<void> seems like a better approach considering the failure mode here, will leave that to the maintainers unless they suggest I change that here).

Additional information

This PR (and all other PRs I'll file in this space) use a recent from-source build of OpenVINO.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: I threw an AI at this, but I have a solid understanding of the mechanics at play and the fix is targeted enough to not cause collateral issues.

[cavusmustafa]

LGTM