This fix bugs on the CPU in fast_compile introduced in:
commit 4f8f0da5
Date:   Thu Feb 27 09:56:12 2014 -0500

Add tile in the doc.

Update the c_code_cache_version that was forgotten.

Mixed

Speed up matVecModM function.

Add consider_constant op

Use pip --user instead of --install-option

I don't know if there is a specific reason that --install-option='--prefix=~/.local' is used instead of pip's --user, but it seems to break things when installing Theano without root access (both myself and Jae Hyun had this issue).

Moved consider_constant from theano.tensor.extra_ops to theano.gradient, moved the optimization that removes it to theano.tensor.opt. Modified theano.compile.builders so it doesn't import theano.gradient, because theano.gradient now imports ViewOp from theano.compile and this would lead to a circular dependency.

Add new subtensor helper make_constant and use it to fix a bug in local_subtensor_merge.

Fix gpu crash and faster optimization

Add fast_compile tag so the optimization to remove the Op always runs. Use g_out.zeros_like instead of tensor.zeros_like so the op also works for sparse variables.

We take care to still have this order: gpu, cpu/gpu fusion, then gpu again.

fix crash when we selected the wrong version of the gpu conv code.

added consider_constant op, which truncates the gradient of an expression to 0. Also added an optimization to remove it from the graph, so other optimizations can be applied properly.

test whether the gradient is computed correctly, and test whether the consider_constant op is removed from the graph.

Update to the blas ops following API change in compyte.