backend: rebase llama.cpp submodule on latest upstream

[cebtenzzre]

This PR updates the llama.cpp submodule to a version rebased on commit ggerganov/llama.cpp@a15ef8f ggerganov/llama.cpp@87e397d.

To rebase successfully:

• Commented-out CUDA logs was replaced with ggml_backend_cuda_log_set_callback
• llama.cpp now only inserts a leading space after BOS, so the hacks in llama.cpp and GPT4All to work around concatenated calls to llama_tokenize were removed

To get it to compile:

• llama.cpp.cmake had to be updated, as the repo was reorganized, and the OpenCL backend was removed
• llama_token_to_piece calls were updated to add the lstrip argument (set to zero)

There were runtime complains about the NONE op not being supported, so I ran test-backend-ops. This crashed partway through, which led to significant changes to the way we are initializing/destroying resources with Kompute. The results of this are:

• We are now running all 243 passing tests instead of only 150, by allowing no-ops on all tensor data types (this also fixes the runtime complaints)
• We now free the device and Vulkan instance as late as possible, as doing it too eagerly can result in repeated re-initialization, which eventually causes the crash during test-backend-ops. Resources are freed by calling Kompute's Manager::clear instead, which has been modified to loop until there is nothing left to clean up.
• An old bug where we did not unref the GPU device on allocation failure was fixed. I believe this could have caused unnecessary VRAM usage after OOM, which would only be freed after a successful model load/unload with Kompute.

[cebtenzzre]

New changes:

• Fix potential crash and memory leaks on OOM in Kompute
• Enable GPT4All support for GPT-NeoX, Gemma 2, OpenELM, ChatGLM, and Jais architectures (all with Kompute support)
• Also enable Kompute support for StarCoder2, XVERSE, Command R, and OLMo

Gemma 2 still doesn't work perfectly due to familiar issues with special tokens in output causing an assertion failure in UI code (likely an unrecognized EOS token).

[manyoso]

Have tested a bit and can't see anything that breaks