Assorted Python 3 fixes

Add TypedListConstant, this fix an optimization warning.

Enable cudnn by default when avaialable in the new back-end

* use Python 2/3 compatible syntax for print;
* use range instead of xrange (creating a list of 100..1000 ints in Python 2 is not a big deal)

Fix Python3 regression in ``pydotprint``

  In Python3 ``map`` returns a map-object (and not a list). This makes the check
  introduced in #3057 fail with "TypeError: unorderable types: map() < list()"

Fix fallback in BLAS detection

Cap memory preallocation to 98% of the total.

set_subtensor GPU realizations

Added tests to test_basic_ops.py for GpuAdvancedIncSubtensor1 (general case) and GpuAdvancedIncSubtensor1_dev20 (2d case if compute capability >=2.0).

Small

add doc test in yaml

Adds test for uint datatype for image_shape and filter_shape in thean…

Gpuarray prealloc

Adds test for uint datatype for image_shape and filter_shape in theano.tensor.nnet.conv and also checks if datatype is in tensor.discrete_dtypes

Multi gpu doc

Add the gpuarray.preallocate configuration variable.  This will help fix out of memory errors in certain sequences of tests.

Rebased cudnn v4

the printout of the context map.