plus some minor code speedup in gradient.py

major changes:

- The self._grad_op now only returns zeros_like() for special types like
    NullType() or DisconnectedType()
- call to grad() will furthur replace returned zero tensors with special
    types
- proposed gradient override interface :
    (single or list of below)
    Ellipsis -> <no_override>
        (-) since python 2 does not support `[...]` syntax, this may
        result in uglier code in python 2
    None -> NullType()
    int(0) -> DisconnectedType()
    OpFromGraph instance or callable -> <override>

minor changes:

- various typo/bug fix

notes:

- This commit breaks OpFromGraph.R_op, which is expected to be fixed in
upcoming commits.

also removed inline test case for infer_shape since the Op is removed
during opt

- much prettier html version

- modified TODO section following Fred's comment

- Moved old docstring in OpFromGraph to function op_from_graph
- Added an new example for op_from_graph docstring
- now we can access `theano.op_from_graph` and
  `theano.OpFromGraph<Inline|Precompiled>`

in `unittesttools.py`:
```
if check_topo:
    topo_shape = shapes_function.maker.fgraph.toposort()
    assert not any(isinstance(t.op, cls) for t in topo_shape)
topo_out = outputs_function.maker.fgraph.toposort()
assert any(isinstance(t.op, cls) for t in topo_out)
```
This seems to check existence of op in `topo_out` regardless `check_topo`.
Since OpFromGraphInline will be replaced with its internal graph at
compile time, this check will always cause error. Thus I moved them
inside the `if` clause.

Update "import numpy" to "import numpy as np" (theano/misc)

Update "import numpy" to "import numpy as np" (theano/compile directory)