Build the runtime support for reductions on NVIDIA GPUs only for `CHP…

…L_GPU_ARCH` (chapel-lang#24006)

Resolves chapel-lang#23893
Resolves Cray/chapel-private#5712

This PR aligns `nvidia` reduction support to `amd`:
https://github.com/chapel-lang/chapel/pull/23950/files#diff-eb05ab5f4c9820ed534cd67b2f2f83426c7fd52e90c86b5afcbb29dc7992296dR24
where we build the reduction support source only for `CHPL_GPU_ARCH`.
This improves the build time of the reduction support by about 4x on my
workstation (1 vs 4 minutes).

Related: chapel-lang#23960

[Reviewed by @jabraham17]

Test:
- [x] nvidia

runtime/src/gpu/nvidia/Makefile.share

@@ -28,10 +28,7 @@ GPU_OBJS = $(addprefix $(GPU_OBJDIR)/,$(addsuffix .o,$(basename $(GPU_SRCS))))
 # a comma-separated list.
 RUNTIME_CXXFLAGS += -x cuda -Wno-unknown-cuda-version \
                     -Xclang -fcuda-allow-variadic-functions \
-                    --offload-arch=sm_60 \
-                    --offload-arch=sm_61 \
-                    --offload-arch=sm_70 \
-                    --offload-arch=sm_75
+                    --offload-arch=$(CHPL_MAKE_GPU_ARCH)
 
 $(RUNTIME_OBJ_DIR)/gpu-nvidia-reduce.o: gpu-nvidia-reduce.cc \
                                          $(RUNTIME_OBJ_DIR_STAMP)