Revert FP16 headers from CUDA 12.2.0 to CUDA 12.1.1

[kmaehashi]

Follows-up #7748. FP16 headers from CUDA 12.2.0 added dependency to vector_types.h, vector_functions.h, cuda_runtime_api.h, etc., causing compilation errors in environment without these headers.

CUDA 12.1.1:

% grep '#include' /usr/local/cuda-12.1.1/include/cuda_fp16.h
/usr/local/cuda-12.1.1/include/cuda_fp16.h:#include "cuda_fp16.hpp"

CUDA 12.2.0:

% grep '#include' /usr/local/cuda-12.2.0/include/cuda_fp16.h
/usr/local/cuda-12.2.0/include/cuda_fp16.h:#include "vector_types.h"
/usr/local/cuda-12.2.0/include/cuda_fp16.h:#include "vector_functions.h"
/usr/local/cuda-12.2.0/include/cuda_fp16.h:#include "cuda_fp16.hpp"

% grep '#include' /usr/local/cuda-12.2.0/include/vector_types.h
#include "crt/host_defines.h"

% grep '#include' /usr/local/cuda-12.2.0/include/vector_functions.h
#include "cuda_runtime_api.h"
#include "vector_functions.hpp"

[kmaehashi]

/test mini

[kmaehashi]

This PR failed on CUDA 12.2 for Windows:

c:\program files\nvidia gpu computing toolkit\cuda\v12.2\include\cuda_bf16.hpp(575): error: no instance of overloaded function "__half::__half" matches the specified type

[leofang]

I'd suggest to merge & revert for now.

[kmaehashi]

For record: no difference in FP16 headers between 12.2.0 and 12.2.1.

[kmaehashi]

/test mini

[kmaehashi]

Tests passed except for CUDA 12.2 on Windows. It seems CUDA 12.2 for Linux is not affected.

[kmaehashi]

Let's proceed to merge. I will:

• Update the doc to say that CUDA 12.2 for Windows is not supported yet, in a separate PR (Add notes for CUDA 12.2 on Windows support #7777)
• Remove cupy.win.cuda122 from required CI (done, Make CuPy run without headers under CUDA 12.2 for Windows #7776)