Add parquet range-read I/O benchmarks (local + S3) (#165)

Two standalone CLI benchmarks that isolate the raw column-chunk range-read cost (no parquet decode): hybrid_scan computes the byte ranges a column projection touches, then only the reads of those ranges into host/device memory are timed.

- parquet_io_benchmark: io_context (vector-io / device_read_async) vs the cucascade::io::datasource per-range host_read/device_read_async, over uring.
- parquet_s3_io_benchmark: native REST (vector-io host / host->device staging) vs kvikIO RemoteHandle datasource, host/device dest.
- CMake: cucascade_parquet_io_benchmark (CUCASCADE_BUILD_IO) and cucascade_s3_parquet_io_benchmark (CUCASCADE_BUILD_S3_BENCHMARK).

Authors:
  - Amin Aramoon (https://github.com/aminaramoon)

Approvers:
  - Bobbi Winema Yogatama (https://github.com/bwyogatama)
  - https://github.com/felipeblazing

URL: #165

Author: aminaramoon

benchmark/CMakeLists.txt

@@ -98,6 +98,20 @@ if(CUCASCADE_BUILD_IO)
 
   target_link_libraries(cucascade_parquet_benchmark
                         PRIVATE cucascade_cudf cucascade_io CUDA::cudart_static)
+
+  # Range-read I/O benchmark: isolates raw column-chunk range reads (io_context
+  # vector-io vs cudf datasource) into host/device, without parquet decode.
+  add_executable(cucascade_parquet_io_benchmark
+                 ${CMAKE_CURRENT_SOURCE_DIR}/parquet_io_benchmark.cpp)
+
+  target_compile_options(
+    cucascade_parquet_io_benchmark
+    PRIVATE
+      $<$<COMPILE_LANGUAGE:CXX>:${CUCASCADE_CXX_WARNING_FLAGS};-Wno-null-dereference>
+      ${CUCASCADE_CUDA_WARNING_FLAGS})
+
+  target_link_libraries(cucascade_parquet_io_benchmark
+                        PRIVATE cucascade_cudf cucascade_io CUDA::cudart_static)
 endif()
 
 # =============================================================================
@@ -123,4 +137,20 @@ if(CUCASCADE_BUILD_S3_BENCHMARK)
     cucascade_s3_parquet_benchmark
     PRIVATE cucascade_cudf cucascade_io kvikio::kvikio ${AWSSDK_LINK_LIBRARIES}
             CUDA::cudart_static)
+
+  # S3 range-read I/O benchmark: raw column-chunk range reads (rest vector-io /
+  # host->device vs kvikIO datasource) into host/device, without parquet decode.
+  add_executable(cucascade_s3_parquet_io_benchmark
+                 ${CMAKE_CURRENT_SOURCE_DIR}/parquet_s3_io_benchmark.cpp)
+
+  target_compile_options(
+    cucascade_s3_parquet_io_benchmark
+    PRIVATE
+      $<$<COMPILE_LANGUAGE:CXX>:${CUCASCADE_CXX_WARNING_FLAGS};-Wno-null-dereference>
+      ${CUCASCADE_CUDA_WARNING_FLAGS})
+
+  target_link_libraries(
+    cucascade_s3_parquet_io_benchmark
+    PRIVATE cucascade_cudf cucascade_io kvikio::kvikio ${AWSSDK_LINK_LIBRARIES}
+            CUDA::cudart_static)
 endif()