Make a version of the code clean.
better test of dx,dy

set the mode in test_opt_unpack because testing an optimization makes no sense when the optimization is not applied

kernels. Modified the list of kernel and image shapes to test in test_conv.py so
that test_ConvOpGrad runs under 2 mins.

Now ConvOp.grad always execute unrolled if the option ask him too do so. This was working correctly when the hash code was not comparing the unroll level. When it started to do so, it was unrolling only for a part of the grad.

some small correction to my last modif on the ConvOp.grad test. Now the test in float32 work BUT with a tolerance of 0.16!

partial rewrite of the ConvOp grad test. It was not testing in float32 correctly. This uncover a bug in the grad in float32. Maybe the bug is in verify_grad. I added an option to verify_grad to help test the float32 case correctly, but we did't can use it by default because it make some test fail. I don't know why.

added an option to verify_grad(cast_to_output_type) to force the casting of the output to the same type as the op expect. This is needed to test the ConvOp grad in float32 type. Otherwise it upcast it to float64.

modified the ConvOp hash calculation. If we don't consider the unroll_* option, we won't be able to correctly test the op as they will be merged together and we will test only the first version. This don't uncover error.

If the env variable THEANO_BASE_COMPILEDIR, we will use it instead of ~/.theano as the base compile directory. We will create a directory based on the os in this directory and put their all file related to the compilation.

This give ~4.5% speed up for unrolled version. Less for other version.