Fix insertion of -arch=... when compiling with nvcc

It used to be inserted only for the first Op after GPU initialization.

Fix insertion of -arch=... when compiling with nvcc
e9d83000 · Pascal Lamblin · cc4fc100 · e9d83000
--- a/theano/sandbox/cuda/nvcc_compiler.py
+++ b/theano/sandbox/cuda/nvcc_compiler.py