fix(cudnn): implement cuDNN 9 error codes, replace todo!() panic with proper mapping

[CharryWu]

Summary

Fixes a runtime panic in crates/cudnn/src/error.rs when the cudnn crate is built against cuDNN 9+.

Problem

When DEP_CUDNN_VERSION >= 90000, the build script sets cfg(cudnn9). The into_result() implementation had an unresolved TODO comment followed by _ => todo!(). Any unrecognized status code would panic at runtime rather than return an error.

cuDNN 9 restructured cudnnStatus_t into a hierarchical numeric system. New sub-codes such as CUDNN_STATUS_BAD_PARAM_NULL_POINTER (2002) and CUDNN_STATUS_NOT_SUPPORTED_ARCH_MISMATCH (3007) could trigger the wildcard arm and cause panics.

Changes

• Add four new CudnnError variants behind #[cfg(cudnn9)]: SublibraryVersionMismatch, SerializationVersionMismatch, Deprecated, SublibraryLoadingFailed
• Replace _ => todo!() with a category-based fallback mapping cuDNN 9 sub-codes to their parent category variant via integer division: 2xxx=BAD_PARAM, 3xxx=NOT_SUPPORTED, 4xxx=INTERNAL_ERROR, 5xxx=EXECUTION_FAILED
• Wire new variants in into_raw() for round-trip correctness

Verification

Verified against cudnn_graph.h from cuDNN 9.20 (CUDA 13.2 / anaconda). The cudnn crate compiles cleanly on this setup.

Testing

• cudnn crate compiles with cuDNN 9.20 on Windows 11 / CUDA 13.2
• No todo!() remains in error handling path
• All cuDNN 8 code paths preserved under #[cfg(not(cudnn9))]