add __cuda as a run-depence for GPU-build

While the GPU build would be runnable on a CPU-only system,
it pointlessly downloads huge binaries, which we can avoid.

recipe/meta.yaml

@@ -193,6 +193,8 @@ outputs:
         # XXX: actually, TF 2.7.0 still demands tb 2.6
         - tensorboard >=2.6,<2.7
         - keras >=2.7,<2.8
+        # avoid that people without GPUs needlessly download ~0.5-1GB
+        - __cuda  # [cuda_compiler_version != "None"]
     test:
       files:
         - test_tensorflow.py
@@ -290,7 +292,9 @@ outputs:
         - snappy
         - sqlite
         - zlib
-
+      run:
+        # avoid that people without GPUs needlessly download ~200-300MB
+        - __cuda  # [cuda_compiler_version != "None"]
     test:
       files:
         - test_libtensorflow.sh
@@ -332,6 +336,9 @@ outputs:
         - snappy
         - sqlite
         - zlib
+      run:
+        # avoid that people without GPUs needlessly download ~200-300MB
+        - __cuda  # [cuda_compiler_version != "None"]
     test:
       files:
         - test_libtensorflow_cc.sh