Cycles: Fix inconsistent command line used for runtime kernel compilation

Basically build-time compiled kernels were using --fast-math (which is correct)
but run-time compiled did not.

intern/cycles/device/device_cuda.cpp

@@ -276,7 +276,8 @@ public:
 		path_create_directories(cubin);
 
 		string command = string_printf("\"%s\" -arch=sm_%d%d -m%d --cubin \"%s\" "
-			"-o \"%s\" --ptxas-options=\"-v\" -I\"%s\" -DNVCC -D__KERNEL_CUDA_VERSION__=%d",
+			"-o \"%s\" --ptxas-options=\"-v\" --use_fast_math -I\"%s\" "
+			"-DNVCC -D__KERNEL_CUDA_VERSION__=%d",
 			nvcc, major, minor, machine, kernel.c_str(), cubin.c_str(), include.c_str(), cuda_version);
 		
 		if(experimental)