Optimized time_substitution method

[vincepandolfo]

Changed the time_substitutions method to build a dictionary of changes to be made by Sympy.subs before calling it instead of calling subs every time as suggested by @mlange05

[mlange05]

Could we please use SymPy's built-in generator traversals sympy.preorder_traversal instead of actually recursing? That should make things even faster.

[mlange05]

Yes, this looks convincing. Can we get an indication of how much faster this is? I would like @navjotk to have a second look at this, but otherwise I'm happy.

[vincepandolfo]

On my machine @mloubout example (https://github.com/opesci/devito/blob/Zhang_tti/examples/tti_example2D.py) is ~6 times faster with this change

[FabioLuporini]

looks reasonable to me

[navjotk]

The change looks good on visual inspection. However, this bit of the code is not currently being tested (I believe). I think now is as good a time as any to add a test for this.

Here's my suggestion for the test: Run a forward propagation with save=True and another with save=False. A comparison of the last time step in both cases should be the assert for the test. Such a test would be able to pick up bugs in this bit of code.

[FabioLuporini]

The test you asked is landing with #67 . Once that PR lands, are you happy with merging this @navjotk ?

[navjotk]

Yes. Now that we have the test in place, we are good to go on this. Merging.