llama.cpp b9990: Minimax2 EAGLE3 support and CUDA 13 builds

llama.cpp b9990 adds EAGLE3 speculative decoding for Minimax2 models — and ships a nullptr fix for the same feature in the same release.

Prebuilt binaries for this release:

• macOS: Apple Silicon (arm64), Intel (x64), iOS XCFramework
• Linux: Ubuntu x64/arm64/s390x CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16
• Windows: x64/arm64 CPU, CUDA 12.4 + CUDA 13.3 DLLs, Vulkan, OpenVINO, SYCL, HIP
• Android: arm64 CPU
• openEuler: x86 + aarch64 for Ascend 310p and 910b with ACL Graph

EAGLE3 is speculative decoding — a draft head predicts tokens ahead, the main model verifies in a single forward pass. For Minimax2, that means higher tokens-per-second on local hardware without changing model weights. The nullptr fix prevents crashes on certain configs, so you don’t need to cherry-pick between commits.

The openEuler builds target Huawei Ascend NPUs (310p and 910b) with ACL Graph mode. That’s a niche but real path for developers running inference on Ascend hardware without leaving the llama.cpp toolchain.

CUDA 13.3 DLLs ship alongside the CUDA 12.4 build — Windows x64 users on newer NVIDIA drivers now have a prebuilt CUDA 13 option without compiling from source.