feat: update llama.cpp to fb34fc262

[abetlen]

Update vendored llama.cpp from v0.4.0 to upstream master at fb34fc262.

• Include the merged Mamba/AMX crash fix from fix(mamba) : make time-step projection input contiguous ggml-org/llama.cpp#28832.
• Bind the FILE-pointer LoRA loader and test-tokenizer enum, and synchronize the sampler-count return type and model-loading comments.
• Refresh the existing Unreleased changelog entry.

Validation: CPU/MTMD build, all 26 tests, C/ctypes ABI and binding smoke checks, Qwen 0.8B image-input smoke test, and Ruff checks passed locally.