Because tests were moved, the RNG for subsequent tests was altered, which
required a change to `utt.assert_allclose` in `tests.tensor.test_basic` for more
reasonable `rtol` settings.

These changes include the addition of `__all__` to `theano.tensor.basic`,
`theano.tensor.io`, `theano.tensor.subtensor`, and `theano.tensor.type_other`.

This change required a lot of general import refactoring in order to avoid (and
remove) circular dependency issues.

One major change involved the removal of imports like `from theano import
tensor`, because the module `tensor` clashes with the type `tensor` (e.g. `x =
tensor()`).  This usage was replaced with the conventional `import theano.tensor
as tt` or `from theano import tensor as tt` (the latter addresses a Python 3.6
import inconsistency/bug).  The same goes for `from theano import scalar`; the
name clashes with the `scalar` type name, so it was replaced with `import
theano.scalar as ts`.

This removes the kludgy object re-definitions that were used to avoid circular
import errors.

These prefixes -- applied to warnings where optimization is impossible because there is no C implementation -- are required to enable the logger to filter out redundant warnings.

Closes #270

When `Scan`'s Cython code is compiled, it will now only hold a lock for its
compilation directory and not unnecessarily block other Theano C compilation
processes.

This change makes `Op.perform` a mandatory method.

Since more than a few `Op`s do not have Python implementations, they've been
made to extend `_NoPython*Op` classes that provide an `Op.perform` that simply
raises a `NotImplementedError`.

This causes `abstractmethod`-decorated methods to take effect (i.e. classes that
don't implement these required methods will fail during construction).

This commit also includes numerous indirect-to-direct `theano.gof` import
corrections.