The test value computation steps were also extracted from `PureOp.__call__` and
put into a stand-alone `compute_test_value` function.

Refactor test value framework so that test value validation is performed up-front.

This test simply loaded an old example of pickled data.  It was both extremely
indirect and brittle.  Changes to the code base will eventually invalidate
nearly all of the old classes used, and whatever the exact issue was, it
should've been tested directly.

The temporary values created during subtensor broadcast-value computations would
cause the test value computation steps to complain/err.  This change prevents
that.

This change also fixes the `set_subtensor` boolean gradient bug in #105.

This change prevents downstream code (e.g. test value computation) from throwing
exceptions (e.g. `ValueError`s) that break the logic in
`theano.tensor.basic.as_tensor_variable`.

- Fix `*Subtensor*` `Op` shape inference
- Make `theano.tensor.basic.as_tensor_variable` consistent in how it handles constant conversions
- Remove `TensorConstant` caching

Now, when `as_tensor_variable` is given `Constant` types, it will return the
same result as a call with the underlying non-`Variable` data.

Closes #98

Closes #99