llama.cpp b9982 fixes per-request reasoning budget overrides in chat completions

The llama.cpp project released version b9982, which addresses a bug in the `oaicompat_chat_params_parse` function where per-request values for `reasoning_budget_tokens` and `reasoning_budget_message` were being ignored.

• The server previously only read server-level defaults for these fields, causing caller-supplied overrides to be silently discarded because the keys were already present in the parameters object.
• The fix ensures that values from the request body are read first, allowing users to explicitly control reasoning budgets (e.g., setting `reasoning_budget_tokens` to 0) or custom messages.
• Unit tests were added to verify that per-request overrides for both tokens and messages are correctly passed to the sampling layer when using a Qwen3 template.
• The release includes binaries for macOS, Linux, Android, Windows, and openEuler across various hardware backends including CPU, CUDA, ROCm, Vulkan, and OpenVINO.