Merge feature (rebased)

Replacement for equivalent nodes and constants are now queued when
the graph is modified, through the MergeFeature, and processed when the
optimization is applied.
This means that appling the MergeOptimization when there is nothing
to be merged is now almost instantaneous, even for big graphs.
It is also faster when lots of nodes have to be merged (empirical 40 %
speed-up for a graph with ~ 5000 nodes).

Default argument fix

Gpu conv faster

Add MergeOptimizer in canonicalize phase

config.blah should not be used as a default argument as the values can
change during execution, but default arguments are only evaluated at
function definition time. The correct idiom is to use a sentinel value
such as None and then check for it, and read out the config value in the
function body.

All the boiller plate code to allow using the run time maximum number of theads on the dimensions 0 of gpu block of threads.

Fix advanced indexing with constants (not variables)

Closes gh-607

fix test in float32, but also check float64 at the same time.