Pass -gpu= at link time for the Thrust extension (#17)

The device-link step is where the final SASS is generated, so an arch passed
only at compile time is discarded and the extension silently gets nvc++'s own
default target instead of what was requested.

Measured on NVHPC 26.1 (aarch64), compile with -gpu=cc90:
  .so linked without -gpu=  ->  sm_100   (the host default, not cc90)
  .so linked with    -gpu=  ->  sm_90

_core_gpu_omp_offload already passes -gpu= at link, which is why only the Thrust
backend was affected.

Author: hfwen0502

setup.py

@@ -397,7 +397,12 @@ def find_nvidia_hpc_sdk():
         # target missing), causing the link to fail with "cannot find
         # -lcublasmp" etc. SBD's GPU path only needs the CUDA runtime, so
         # explicitly link -lcudart instead.
-        extra_link_args=extra_link_args + ['-mp', '-cuda', '-lcudart'],
+        #
+        # -gpu= is repeated at link because the device-link step generates the
+        # final SASS: without it nvc++ silently targets its own default instead
+        # of the requested arch(es).
+        extra_link_args=extra_link_args + ['-mp', '-cuda', f'-gpu={gpu_arch}',
+                                           '-lcudart'],
     )
     ext_modules.append(gpu_thrust_ext)