Keep deferred copy count storage alive

[kevmo314]

Summary

• fix deferred DtoH copy-count writes to use caller-owned storage that stays alive until rpc_write_end()
• restore cuCtxSynchronize stdout capture instead of disabling it
• fixes vectorAdd_nvrtc hanging because the client no longer reads a corrupted deferred-copy count

Root cause

rpc_write() stores iovec pointers and flushes them later in rpc_write_end(). lupine_write_pending_dtoh_copies() queued a pointer to its local count variable, so calls made before rpc_write_end() could reuse that stack slot. In vectorAdd_nvrtc, the client read garbage for copy_count and then waited for nonexistent copy payload bytes.

Verification

• cmake --build build -j"24"
• ctest --test-dir build --output-on-failure
• RESULTS_DIR=/tmp/lupine-vectoradd-count-lifetime SAMPLE_TIMEOUT=30 BUILD_SAMPLES=0 SERVER_PORT_BASE=15180 ./test/run_cuda_samples.sh vectorAdd vectorAddDrv vectorAdd_nvrtc simplePrintf