Reduce packaged native library size (#4881)

## Summary
- rely on the static nvCOMP copy already linked into `libcudf.so`
- stop packaging and loading the redundant nvCOMP shared libraries
- localize symbols from statically linked CUDA compiler and nvCOMP
implementation archives with `--exclude-libs`
- retain the regular ELF symbol table for stack traces and debugging

No GPU architectures are removed.

No upstream cuDF change is required. cuDF already selects
`nvcomp_static` when
`CUDF_BUILD_STATIC_DEPS=FORCE`, as used by this build. This supersedes
the previous dependency on
rapidsai/cudf#23415.

Related to NVIDIA/cudf-spark#15145.

## Size results
Focused CUDA 12.9.1 Release builds with `CMAKE_CUDA_ARCHITECTURES=75`:

| Variant | Packaged native bytes | ZIP level 6 |
| --- | ---: | ---: |
| main: static nvCOMP in `libcudf.so` plus packaged `libnvcomp.so` |
754,765,376 B | 389,662,932 B |
| this PR: static nvCOMP only in `libcudf.so`, private symbols localized
| 693,387,896 B | 354,470,480 B |

The focused build reduces packaged native data by 58.53 MiB uncompressed
(8.13%) and 33.56 MiB
after DEFLATE (9.03%). `libcudfjni.so` is unchanged and omitted from
both rows.

These measurements use one architecture to make repeated comparisons
practical. The implementation
does not change the production architecture list.

## Linking and symbol compatibility
- the JAR contains `libcudf.so` and `libcudfjni.so`, with no nvCOMP
shared-library entries
- `libcudf.so` has no nvCOMP `DT_NEEDED` entry
- all 12 nvCOMP JNI entry points remain dynamically exported
- nvCOMP implementation symbols are absent from `.dynsym`
- the regular ELF `.symtab` remains available for diagnostics

## Validation
- native CUDA 12.9.1 Release build with `CMAKE_CUDA_ARCHITECTURES=75`
- `mvn package -DskipTests` through the normal patch lifecycle
- `NativeDepsLoaderTest`: 4 passed
- GPU `NvcompTest` LZ4 and Zstd cases: 2 passed
- `git diff --check`

---------

Signed-off-by: Gera Shegalov <gshegalov@nvidia.com>

Author: gerashegalov

patches/0001-use-static-nvcomp-from-libcudf.patch

@@ -0,0 +1,14 @@
+diff --git a/java/src/main/java/ai/rapids/cudf/NativeDepsLoader.java b/java/src/main/java/ai/rapids/cudf/NativeDepsLoader.java
+index 0ab9ec79d1..3d42449128 100755
+--- a/java/src/main/java/ai/rapids/cudf/NativeDepsLoader.java
++++ b/java/src/main/java/ai/rapids/cudf/NativeDepsLoader.java
+@@ -69,9 +69,6 @@ public class NativeDepsLoader {
+    * subsequent stages are loaded.
+    */
+   private static final String[][] loadOrder = new String[][]{
+-      new String[]{
+-          "nvcomp"
+-      },
+       new String[]{
+           "cudf"
+       },

patches/noop.patch

@@ -570,14 +570,6 @@
                     <include>libprofilerjni.so</include>
                   </includes>
                 </resource>
-                <resource>
-                  <directory>${libcudfjni.build.path}</directory>
-                  <includes>
-                    <include>libnvcomp.so</include>
-                    <include>libnvcomp_gdeflate.so</include>
-                    <include>libnvcomp_bitcomp.so</include>
-                  </includes>
-                </resource>
               </resources>
             </configuration>
           </execution>

pom.xml

@@ -62,6 +62,7 @@ option(BUILD_TESTS "Configure CMake to build tests" OFF)
 option(BUILD_BENCHMARKS "Configure CMake to build (google) benchmarks" OFF)
 option(BUILD_FAULTINJ "Configure CMake to build fault injection" ON)
 option(BUILD_PROFILER "Configure CMake to build profiler" ON)
+option(HIDE_PRIVATE_CUDA_SYMBOLS "Hide private symbols from static CUDA compiler libraries" ON)
 
 message(
   VERBOSE "SPARK_RAPIDS_JNI: Build with per-thread default stream:
@@ -70,6 +71,9 @@ message(
 message(VERBOSE "SPARK_RAPIDS_JNI: Configure CMake to build tests: ${BUILD_TESTS}")
 message(VERBOSE "SPARK_RAPIDS_JNI: Configure CMake to build (nvbench) benchmarks: ${BUILD_BENCHMARKS}")
 message(VERBOSE "SPARK_RAPIDS_JNI: Configure CMake to build fault injection: ${BUILD_FAULTINJ}")
+message(VERBOSE
+        "SPARK_RAPIDS_JNI: Hide private static CUDA compiler symbols: ${HIDE_PRIVATE_CUDA_SYMBOLS}"
+)
 
 set(SPARK_RAPIDS_JNI_BUILD_TESTS ${BUILD_TESTS})
 set(SPARK_RAPIDS_JNI_BUILD_BENCHMARKS ${BUILD_BENCHMARKS})
@@ -349,6 +353,18 @@ target_link_libraries(
     ${PARQUET_LIB}
     ${THRIFT_LIB}
 )
+if(HIDE_PRIVATE_CUDA_SYMBOLS)
+  # These toolkit implementation symbols are not part of the cuDF or JNI ABI. Localizing them keeps
+  # the regular ELF symbol table for diagnostics while removing them from the dynamic export table.
+  target_link_options(
+    spark_rapids_jni
+    PRIVATE "LINKER:--exclude-libs,libnvrtc_static.a"
+            "LINKER:--exclude-libs,libnvrtc-builtins_static.a"
+            "LINKER:--exclude-libs,libnvptxcompiler_static.a"
+            "LINKER:--exclude-libs,libnvJitLink_static.a"
+            "LINKER:--exclude-libs,libnvcomp_static.a"
+  )
+endif()
 rapids_cuda_set_runtime(spark_rapids_jni USE_STATIC ON)
 set_target_properties(spark_rapids_jni PROPERTIES LINK_LANGUAGE "CXX")
 # For backwards-compatibility with the cudf Java bindings and RAPIDS accelerated UDFs,