cuda.core: validate ctypes host callback signatures against CUhostFn - #2525

Summary

Closes #2439.

cuda.core accepted any ctypes._CFuncPtr as a host callback and cast its address to CUhostFn without checking the declared prototype. CUDA requires void (*)(void*); a mismatched restype, argument list, or calling convention is undefined behavior when the driver invokes the callback.

_resolve_host_callback now validates ctypes prototypes against that ABI before the pointer reaches CUDA. Construction (GraphNode.callback / GraphBuilder.callback) and update paths share the check. Valid CFUNCTYPE(None, c_void_p) callbacks keep working; on Windows WINFUNCTYPE is also accepted. Docs, stubs, and the 1.2.0 release notes are updated accordingly.

Changes

• Reject incompatible ctypes host-callback signatures with TypeError (message names CUhostFn and suggests a conforming declaration or a Python callable)
• Document the required ABI on callback / update APIs
• Add construction, update, and stream-capture rejection tests (plus Windows WINFUNCTYPE accept, skipped on Linux)
• Note the stronger checking in cuda_core/docs/source/release/1.2.0-notes.rst