[BUG]: Caching of load_nvidia_dynamic_lib result implies "never close the handle". #1011

Summary

cuda.pathfinder.load_nvidia_dynamic_lib(libname) returns a LoadedDL whose OS handle (_handle_uint) must not be closed by callers. The function is decorated with functools.cache, so the same handle instance is shared across
the process. If user code calls dlclose (Linux) or FreeLibrary (Windows) on that handle, it can unload the module while other code still uses it, resulting in crashes or subtle failures.

This behavior is by design of the current caching, but it was not explicitly documented until now.

Affected platforms

• Linux and Windows (these are the only supported platforms for this loader).

How it shows up

• After calling load_nvidia_dynamic_lib(...), code looks up symbols via dlsym / GetProcAddress, then unloads the library.

• Later, other code (or a subsequent call to load_nvidia_dynamic_lib) uses the cached handle and crashes or fails to resolve symbols.

Minimal example (illustrative)

from cuda.pathfinder import load_nvidia_dynamic_lib

loaded = load_nvidia_dynamic_lib("cudart")
handle = loaded._handle_uint

# ❌ Do not do this – it can unload a cached, shared handle:
#   Linux: ctypes.CDLL("libdl.so").dlclose(ctypes.c_void_p(handle))
#   Windows: kernel32.FreeLibrary(ctypes.c_void_p(handle))

# Later use of the cached handle may now fail or crash.