This last commit finish to bring from 584s to 66s the slow scan test:
theano/scan_module/tests/test_scan.py:T_Scan.test_hessian_bug_grad_grad_two_scans

The changed optimization:

local_abs_merge, local_mul_switch_sink,
local_upcast_elemwise_constant_inputs,
local_remove_switch_const_con

are called on nodes that aren't yet part of the graph, during
local_subtensor_merge optimization. They aren't supposed to work on
node that are in the graph, but as they called
get_scalar_constant_value() they traversed it very frequently!

This pre_greedy_local_optimizer call is needed as it would introduce
too much scrapt that will be optimized away. So we preoptimize the
graph.

As now we do the constant_folding in the phase
(canonicalize/stabilize/specialize) where those changed opt are used,
they can work directly on Constant. Other opt will make sure their
inputs get constant if possible.

Indent what is inside. I'll use that in the next commit. This will make the change more visible.

Fixed python3 compat issue introduced in 80ebe9c3.

Bytes need to be decoded prior to performing any string operations, such
as split/strip, on them.

Added usecases for scan tutorial

Fix usage of Popen.wait().

Recursion limit

It was making too much tests fail, but it still should keep the opt I wanted.

Speed up. get_scalar_constant_value and slow scan test.

This make less condition being tested. It speed up the slow scan tests.

The 2 subtensors tests change probably due to optimization order changes.
The 2->22 is that in the past, the Composite wasn't flattened, so the outer Composite was smaller.

We fix for atleast 158 loops of that problem.

with Composite.

This allow to pickle bigger Composite graph as it simplify the graph,
so we bust later the max recursion limit.

Flatten Composite.

This help gh-689 as it allow it to support up to 57 iteration of the
loop instead of just 43.

This could also help Theano cache in the case where graph could have
different fusion order, so different way the composite are in each
other, but do the same computation.

This make the printing of Composite more readable.

Fix usage of Popen.wait().  This is liable to deadlock if used with subprocess.PIPE and it was used like that all over the place.

I've created new wrapper that will hopefully avoid this broken usage.

Faster optimizer

Make PatternSub use that feature to lower optimizer time by tracking
less frequent node. This allow to save 85s on 317s in optimizer time.

Mixed

Faster opt (and other unrelated things).

Cuda fortran order

Add other gpualloc tests. It currently fail as ger isn't implement on
the new back-end.