Support MTML 2.4.1 with legacy runtime fallback

[yeahdongcn]

Summary

• prefer the versioned Linux library libmtml.so.2, with libmtml.so as a legacy fallback
• align ctypes structures and wrappers with the MTML 2.x C ABI while preserving compatibility aliases
• bump the Python package version to 2.4.1 to match the bundled MTML runtime

Validation

• python -m pytest -q test_abi_layout.py test_library_loading.py (13 passed)
• C11 sizeof/offsetof ABI probe against the MTML header
• S80, MTML 2.4.1 and legacy MTML 2.2.0: core API smoke plus test_pymtml.py, test_pynvml.py, and test_sglang_compat.py
• S5000, MTML 2.4.1 and legacy MTML 2.2.0: the same API/test matrix
• built and installed the mthreads_ml_py-2.4.1-py3-none-any.whl artifact in an isolated environment

The MTML 2.4.1 library on S5000 reports Not Supported for several max-clock queries that returned values with MTML 2.2.0; the binding propagates that library result correctly.

No active codec sessions, MPC partitions, virtualization instances, or retired ECC pages were available. Their non-empty payload layouts are covered by the C-header ABI checks rather than live device samples; the target runtime image does not include a separate MTML 2.4.1 development header.