Prevent undefined behavior when passing handle from Treelite to cuML FIL

[hcho3]

Problem. FIL unit tests have been failing when run with the pip wheel:

FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-1-native-2-False-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-1-native-2-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-1-native-10-False-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-1-native-10-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-1-native-20-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-10-native-2-False-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-10-native-2-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-10-native-10-False-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-10-native-10-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-GradientBoostingClassifier-10-native-20-True-30-1000] - MemoryError: std::bad_alloc
FAILED test_fil.py::test_fil_skl_classification[2-RandomForestClassifier-1-native-2-False-30-1000] - MemoryError: std::bad_alloc

In the meanwhile, the same tests were passing when run with the Conda package.

Diagnosis. When a Treelite model object is passed from the treelite Python package into the cuml Python package, std::string fields get corrupted. This is likely caused by C++ ABI incompatibilities, as different C++ compiler toolchain/version etc are used when building cuML pip wheel and Treelite pip wheel. More precisely, libtreelite in cuml is incompatible with libtreelite in treelite, and it's unsafe to directly pass a handle from one to another.

Solution. Have the treelite Python package export a serialized byte sequence, and have cuML re-construct the Treelite model from the bytes. TL2cgen uses this technique to enable loose coupling between itself and Treelite.

[jakirkham]

Thanks Philip! 🙏

Think we might be able to type these as well. Though please double check the types in case I missed something

[hcho3]

I will need to make the same fix for the experimental FIL.

[jakirkham]

Thanks Philip! 🙏

Think the error message could be formatted a bit nicer. Made a few minor suggestions related to this. Not a big deal either way, but maybe this is helpful

[hcho3]

Todo. Check for possible memory leak.

[bdice]

Thanks for this fix @hcho3!

[hcho3]

/merge