Avoid recursing into already substituted expressions

[fjetter]

Closes #835

This should avoid the worst case recursion pattern I described in #835

For the TPCH query this indeed reduces the number of substitute invocations (beyond the guard) from ~13M down to 17k

[fjetter]

We are avoiding the runtime complexity problem with this change but for some queries the substitute operation is still the slowest part. During substitution there are many new objects created (even with #798) so many new _names have to be computed which is often a quite expensive operation

[phofl]

thx! This is what I meant with caching in the issue