Add test for gh-1549

Fix scan merge inouts

We want to avoid merging different outputs of a scan node, that compute
the same thing, but keep a different number of steps.

In particular, use op.__call__ rather than make_node, so the
tag.test_value gets computed automatically.

The weird syntax with **dict(return_list=True) is because in older
Python versions (including 2.4), you cannot pass a named argument
(return_list=True) after an expanded list of unnamed arguments (*args).
You can, however, pass an expanded dictionary of named argument
(**kwargs).

Add acknowledgement about the MRG31k3p implementation.

Fix pickling of GpuConv 

Minor typo fixes and rephrasing for clarity

fix test following the removing of set_flops.

Flops