We should not try to access C array before index 0!

On a 32-bit machine, numpy.take may refuse to use int64 arrays for indices.
(incl. typo fix by @dwf)

Doc comment

Blas opt fix

Traceback comments

Removed keep_wrt_type parameter of tensor.grad

The reason I had added this parameter in 0b8b021c was to let the user get rid of
annoying warnings. However, I probably shouldn't have given the option to keep
the old behavior with no warning, since this behavior was unintuitive and
potentially dangerous. Since the warning was removed by 965afb68, it seems
pointless to keep this extra parameter that adds extra complexity to the
tensor.grad interface with no benefit.

Import fix

- Reordered imports in a more logical way
- Uniformized notations to use tensor.* everywhere instead of T.*, basic.*,
  theano.tensor.* and theano.tensor.basic.*

changed default value of keep_type in accordance with decision from
theano-dev mailing list
added unit test to ensure that grad behaves as it is required to for
pylearn2 support

Got rid of a few "except:"

'except Exception' is to avoid catching interrupts like Ctrl-C.

A few fixes to prepare for change in grad output

I caught places where the warning about using grad w.r.t. a one-element
list was raised when running the test-suite.
This commit removes these warnings (ensuring changing the default value of
keep_wrt_type to True will not affect tests).

Note that I changed the behavior of 'test_dot_w_self'. I asked the
author (James) before doing so: it turns out it was not behaving as
intended before.

Docfix

__hash__ should take only 1 argument.

Added new option 'keep_wrt_type' to tensor.grad

This allows one to decide which behavior is desired (future new behavior
or old one) and get rid of the warning at the same time.

Print the name of the optimizer when we have it.

We disable amdlibm for nvcc. This was previously done in half.