The 'remove' method was bugged (it would not remove an item if it was the last
item in the queue).
This commit actually replaces the full custom deque implementation by a subclass
of the original Python 2.4 implementation of deque, as it seems safer than
re-writing everything.

The implementation is also moved to the python25.py file since it is the place
for Python 2.4 compatibility code.

Test using None in slices in local_subtensor_merge

Fix ticket number (oops).

Replace refs to Trac tickets with Assembla

Update dev start guide

Update the online doc about max, min,... function.

Finished square diagonal op

Check blas improvements

docstrings for the pinv op

Improved overall presentation

Fix subtensor merge

The formula for getting an element from an index in a reversed sequence
was wrong.

- Simplify and move test case for the bug reported by Razvan
- Add more general test of merging tensor[slice][index]

I think it is a bit more clearer as documentation compared to the text I've
wrote.

Scan revert fix

The problem was that CudaNdarrays would not allow you to zero some part of
them. My fix was to allocate a ndarray and have then the CudaNdarray setitem
to copy that memory on the card. Fred fixed the bug on the CudaNdarray side
though so this fix (which is suboptimal) is not needed anymore.

Sandbox pseudo-inverse

transpose method and tests