failed with floatX=float32.

* gradient now supports non-diagonal inputs
* ExtractDiag won't crash if input has some 0-dimensions

Using new variable comparison mechanism in tests

fix test in FAST_COMPILE mode. There is no perform method defined.

example of how scan can terminate on condition

Moved the developer guide to the main theano web page.

Windows: Moved the Theano compile directory to the local part of the user profile
to prevent roaming of temporary files.

Arrays of length 0 are forbidden by C standard and treated as errors by the
Visual C++ compiler.

Applied patch
from https://groups.google.com/d/topic/theano-dev/CCJXDUOaaUE/discussion

Moved function implementations from header theano/sandbox/cuda/cuda_ndarray.cuh
to theano/sandbox/cuda/cuda_ndarray.cu. This was necessary to fix a crash
in free() on Windows caused by two created DLLs exporting the same symbols
for the code in the header file.

Tested using CUDA 4.0 and CUDA 3.2 with Visual Studio 2010 and 2008 on Windows 7.

Fixed optimization for exp(x) * sigmoid(-x)

Merge branch 'documentation_equal_computation' of https://github.com/pascanur/Theano into pascanur-documentation_equal_computation

Conflicts:
	theano/scan_module/scan_utils.py

rebased Razvan's Rop/Lop pull request
It was checked by me and Ian.

New user-friendly function to compare variables

Great commit. Everything looks nice. We should make a ticket though that at one point in time we should remove one of the two implementations .. 

Fix test

Minor