Execute as "python -m pegen.vm_generator" in the Tools folder

…envm

This is not complete, but I want to get to left-recursion before I fix
this, and I don't actually understand the code generator well enough
to know how to make it work for `&('foo'|'bar')`.

It definitely makes parsing xxl.py 5-10% slower. :-(

This was super easy.

Benchmark time for xxl.py is now around 2.030 seconds.
(But this is with a super simple grammar. We'll have to see what is will be with the real grammar.)

Also silence compiler warning about default case in call_action().

Also cleaned up the type of the second arg to add_opcode() -- it now
must always be a string, the one call that didn't pass a string now
calls str().

(The bad news: it's currently twice as slow as the 'new' parser.)

This reduces the number of is_memoized calls dramatically, to what it
is for the recursive-descent parser (+1 for the root).

However we still have 50% more calls to insert_memo.  This has to be
investigated later.

It's not any faster than before though.