Micro optimization indexing using tuples instead of lists

Since Python 3.X (i can't remember which) tuple concatenation like this
is fast since python inspects the tuple usage and knows that it is the
last reference and reuses the underlying c object.

The change to LazilyIndexedArray is that `shape` is repeatidely accessed
throughout the codebase (ndim, and shape are heavily used) and thus we
benefit from pre-computing this at creation time.