[BUG]: CUB is missing when installing on CUDA 10.2

### 🐛 Describe the bug

Even though our docs say we require CUDA 11 and above, some users may still use CUDA 10.2. As `cub.h` is missing in CUDA 10.2 built-in, we have to provide extra documentation to guide users to use ColossalAI on CUDA 10.2 and support comaptibility test.

### Environment

_No response_