feat: add public runtime dispatch API

[voltjia]

Summary

Adds a stable public dynamic dispatch layer for InfiniRT runtime calls, so users can call CUDA-style free functions through #include <infini/rt.h> instead of selecting Runtime<Device::Type::...> directly.

Changes

• Add public runtime APIs: SetDevice, GetDevice, GetDeviceCount, DeviceSynchronize, Malloc, Free, Memset, and Memcpy.
• Extend public header generation to also generate generated/src/runtime_dispatch.cc for enabled backends.
• Add backend runtime adapters for CPU, NVIDIA, Iluvatar, MetaX, Moore, Cambricon, and Ascend device management and memcpy kinds.
• Add CTest coverage for CPU/NVIDIA dynamic dispatch and extend install consumer smoke to use only <infini/rt.h>.

Validation

• ruff check scripts/generate_public_headers.py
• ruff format --check scripts/generate_public_headers.py
• git diff --check origin/master..HEAD
• Remote CPU CTest: 6/6 passed
• Remote NVIDIA CTest: 6/6 passed
• Remote Cambricon CTest: 4/4 passed
• Remote Iluvatar CTest: 4/4 passed
• Remote Ascend CTest output: 4/4 passed; container exits with code 137 after tests complete, consistent with the known environment behavior