Fix broadcast pattern of dnn conv op.

Revert change that break too many test in floatX=float32

[CRASH] we where returning a python int while this was reused in Elemwise.perform that expect ndarray.

language fix

Fix scan_args for the 'as_while' case properly.

Fix crashes related to sparse

Make subtensor.make_node infer more broadcastable. This remove crash from local_subtensor_merge that was loosing broadcastable dimensions.

Make sure we remove the exe_path even if try_run is False in GCC_compiler.try_compile_tmp().

Fixed bug with TypedList example

conv2d meta-optimizer missing commits

Do not enforce that if the input is sparse, we need the grad to return
sparse.

I do not force the grad to be the same format. This would require
conversion.  I do not think we always enforce this already, so I do
not want to enforce that now.

Also make theano.grad test the output is good.

Turn off test_value when checking for beta init.

conv2d meta-optimizer

Use clang++ on mac (if present).

Enable again dnn pool grad after fixed it.

bug: The optimization where converting the ignore_border=False to ignore_border=True.

crash: The optimization that convert the gpudownsample grad to
cuddpoolgrad was swapping the inputs parameter. This is caused in part
by cudnn using strange definition of input and output. Now the op use
the same inputs order as the other pool grad ops. The only difference
is tha the c_code() use the name that cudnn use. There is a comment
that tell that.

Add grad() to GpuDnnConvGrad* ops