constructor.
* Check that all instruction ids are unique on kernel creation.
* Add a uniquify_instruction_ids() transform.

This closes #89.

Among other things, passing fixed parameters can improve the accuracy
of temp shape inference, so that fallback can occur less often.

The motivation for this is to remove the temp_shape_fallback warnings
when constructing sumpy kernels.

This changes implements a LazyList and
LazyListWithEqAndPersistentHashing data structure. These data
structures lazily unpickle their contents.

This change also renames LazilyUnpicklingDictionary to LazyDict.

cacheable before creating the cache key.

This function does trivial things, but it's there to solve the problem
that the reduction neutral element getters are not allowed to store
dtypes (#80).

The function mangler demands that a function knows its type based on
its arguments. For the neutral element getters, this is impossible
because they take zero arguments. The simplest fix I can think of is
to change a call to neutral_element() to a call to make_tuple().

Currently, the tuple code doesn't work yet due to pickling issues.  I
think the root cause is somewhere in
__hackily_ensure_multi_argument_functions_are_scoped_private().

get included in dependency calculcations.

Add a save and reload test that uses base_storage.

array prefixes as they are aded to the set.

Closes #25 on gitlab
Depends on inducer/pytools!2 on gitlab

type. Now, the argument to a reduction can be either a scalar or a
tuple.

This reverts commit 7913e68a.

either a tuple, reduction, or a substitution invocation).

Also, rewrite the reduction mapping code to rely on recursive calls
map_tuple(). This cleans up a lot of reduction mapping code
substantially.

* Eliminates inames arguments to reduction functions (closes #32).
* Changes the argmax function to accept two arguments.

See also: #62

and __setstate__() doesn't run if the state is False.

Ensure that we always return a non-false state (another version bump
here).