Reduce packaged native library size

[gerashegalov]

Summary

• link libcudf.so against the libnvcomp.so already packaged by the Java distribution instead of embedding a second static nvCOMP copy
• localize symbols from statically linked CUDA compiler implementation archives with --exclude-libs
• retain an opt-out via HIDE_PRIVATE_CUDA_SYMBOLS=OFF

No GPU architectures are removed.

Dependency

Draft until NVIDIA/cudf#23415 merges and the submodule revision is available from upstream cuDF.

Related to NVIDIA/cudf-spark#15145.

Size results

CUDA 12.9.1 Release builds for sm_75, changing only the two settings measured below:

Variant	libcudf.so	ZIP level 6
static nvCOMP, CUDA symbols exported	711,163,904 B	359,193,871 B
shared nvCOMP, CUDA symbols exported	691,845,960 B	344,989,786 B
shared nvCOMP, CUDA symbols localized	668,512,072 B	336,370,271 B

Combined reduction: 40.68 MiB uncompressed and 21.77 MiB after DEFLATE. The Maven-built JAR stores libnvcomp.so, libcudf.so, and libcudfjni.so as DEFLATED entries.

Symbol compatibility

The final library removes 195,864 private CUDA toolkit symbols from .dynsym, with no added symbols and no changes to JNI or cuDF mangled exports. The regular ELF .symtab is retained.

Focused crash tests retained private function names in HotSpot fatal-error logs and names/source lines in GDB. As expected for localized symbols, glibc backtrace_symbols reports offsets rather than private CUDA implementation names; production sources do not use that API.

Validation

• full native Release build on CUDA 12.9.1 (sm_75)
• loaded libnvcomp.so, libcudf.so, and libcudfjni.so in Java dependency order
• mvn package -DskipTests in the project CUDA container
• NativeDepsLoaderTest: 4 passed
• git diff --check

[gerashegalov]

Superseded by #4881, which adds the required DCO sign-off on a new branch without rewriting published history.