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.