python3Packages.tensorflow-bin: fix cuda routines that require jit

Without this change (or $XLA_FLAGS set properly), errors like this would
crop up, resulting in JIT compilation failures:

2025-10-07 17:29:56.600986: W external/local_xla/xla/service/gpu/llvm_gpu_backend/default/nvptx_libdevice_path.cc:40] Can't find libdevice directory ${CUDA_DIR}/nvvm/libdevice. This may result in compilation or runtime failures, if the program we try to run uses routines from libdevice.

pkgs/development/python-modules/tensorflow/bin.nix

@@ -205,7 +205,12 @@ buildPythonPackage rec {
           ''}
         done
       ''
-    );
+    )
+    # Symlink nvcc besides TensorFlow so that routines that require JIT can work
+    # properly.
+    + lib.optionalString cudaSupport ''
+      ln -s ${cudaPackages.cuda_nvcc} "$out/${python.sitePackages}/tensorflow/cuda"
+    '';
 
   # Upstream has a pip hack that results in bin/tensorboard being in both tensorflow
   # and the propagated input tensorboard, which causes environment collisions.