Note sure it makes the file anymore readable, but at least I've tried.

Otherwise the numberic differentiation gradient quickly becomes numerically
unstable.

the index goes from 0 to n_mit_mot-1 for mit_mot sequnces.

The main bug was gradients where represented as shared variables.
Now we represent them as sit_sot sequences to which only the last step
is used (hence the savemem optimization does the memory clean up). The
advantage is that gradients with respect to sitsot are well defined.

When collecting shapes from the outputs of the inner function, we should
keep track that mit_mot arguments have multiple outputs for one argument.

The error was raised if either the dtype didn't match or ndim didn't match.
However the error message did not display the ndims.

Fixed max/argmax gradient tests in float32

fix how we compute the number of elements in the compile dir.

Work around bug in numpy 1.6 with ufunc.reduce

fix for fails in the buildbot

The tolerance needs to be increased because the numbers in the computations
are too small.

Fixes for Windows and Python 2.4

Fixed import error in a sparse test

Minor stuff (PEP8 and typos mostly)

The rephrasing is to avoid using 'type' for class name, given that
'type' has a special meaning in Theano.

Add "filter_variable" mechanism in Type

Jacobian/Hessian

If one of the "then" elements is a TensorType, and the corresponding
element in the "else" branch is a compatible type, or vice versa,
convert it into TensorType, so both have the same Type, and the
output Type can be determined.

This should fix a problem reported by Ian.
[Commit amended to fix problem spotted by Fred]