Improve file-based kernel cache

• Often ~/.cupy/kernel_cache inode becomes very big. I think it’s better to introduce a hierarchy e.g. using the first 2 chars from the cache key: ~/.cupy/ab/abXXXXXXXXX....cubin.
• Avoid appending MD5 to a file. The reason this is introduced was to support Windows that needs special care to do atomic rename (https://github.com/cupy/cupy/pull/37/files#r115385936). os.replace is in Python 3 so we can just use it now. (FYI, on Windows os.replace uses MoveFileExW with MOVEFILE_REPLACE_EXISTING flag which can be considered atomic in our usecase; there is a long thread in the forum.)
• Include some hint info to filename to manage cache files easily. (maybe NVRTC version and compute capability)

[kmaehashi]

How about adding this tool https://github.com/cupy/cupy/blob/master/.pfnci/trim_cupy_kernel_cache.py to cupyx/tools?

Include some hint info to filename to manage cache files easily. (maybe NVRTC version and compute capability)

Also including a CuPy’s major version (‘8’, ‘9’, etc.) to filename may help managing caches.

[leofang]

related? #1890