Fixed tests when on_shape_error = 'raise'

Scan check

It is expected that some implementations of 'infer_shape' raise a ShapeError to
indicate that they do not know how to compute the shape. In such a situation,
we should silently fall back to the default implementation, regardless of the
value of 'on_shape_error' (whose name may be misleading for developers, since
we want to catch all exceptions *except* ShapeError).

Also, when a bad shape was inferred, we should not raise a ShapeError (which
may be caught silently), but rather something that indicates that there is a
serious bug somewhere, like an AssertionError.

Minor typo fix and more consistent use of rand

inputs.

parameters.

the node.

property

I removed some re-definitions of rand that were pointless (as far as I can
tell).

I verified that tests in test_basic.py still pass in floatX=float32 and
float64.

fix gpu tests when using theano flags device.

Fix an import case and made more test run in floatX

Minor fixes

Minor fixes

This is to be consistent with the idea that "on_shape_error=warn" is supposed
to be a warning, not an error.
NB: I verified that the warning is displayed on screen by default.

Fixes for shape opt

One test was forced to float64, as otherwise the test crash. I make an issue to fix it.

Workaround for optimization failure

Updated NEWS.txt and doc/NEWS.txt

- Removed TODO from NEWS.txt (since it has been updated already)
- Copied NEWS.txt to doc/NEWS.txt (just a basic copy, no need to review that
  part)