[BUG] Stop using private RMM macros in cudf

[bdice]

Describe the bug
cuDF is using private RMM macros and should stop relying on these.

Sources:
https://github.com/search?q=repo%3Arapidsai%2Fcudf+%2F%28%3F-i%29RMM_%5BA-Z_%5D*%5C%28%2F&type=code

The specific examples I see are:

• cudf/cpp/src/interop/to_arrow_device.cu

  Line 481 in 0e01c48

  RMM_ASSERT_CUDA_SUCCESS(cudaEventDestroy(private_data->sync_event));

• cudf/java/src/main/native/src/RmmJni.cpp

  Line 660 in 0e01c48

  RMM_CUDA_TRY_ALLOC(cudaMalloc(&ptr, size), size);

• cudf/java/src/main/native/src/RmmJni.cpp

  Line 672 in 0e01c48

  RMM_ASSERT_CUDA_SUCCESS(cudaFree(cptr));

Pull requests like #18108 should be unnecessary.

Expected behavior
No RMM_... macros are used in the cudf repository.

xref: rapidsai/rmm#1848

[bdice]

@pmattione-nvidia -- could you investigate a fix for the JNI side?

[pmattione-nvidia]

Seems like the easiest thing to do is just to copy/paste the macro body in place at those locations. The thing I don't know is whether RMM has a dedicated interface to do these actual things that I should call instead

[bdice]

@pmattione-nvidia Most macros are duplicated across RAPIDS libraries, to avoid reliance on internal implementation details of other libraries (example: rapidsai/kvikio#653). I would recommend implementing macros named specifically for the cuDF JNI layer.

The macro in to_arrow_device should probably be a cuDF macro instead of an RMM macro.