rpython: simplify and probably remove a performance bottleneck

[asarhaddon]

Traverse environments without recursion. Unwrap the key once. Do not wrap DEBUG-EVAL on each evaluation.
Spare two intermediate MalLists for the 'do' special form.

[kanaka]

Looks good. I've merged them.

Here are some quick tests on my local machine:

• 214,000 iters/sec -> before these changes
• 282,000 iters/sec -> with these changes
• 379,000 iters/sec -> with these but with DEBUG-EVAL commented out

Current Java implementation is 274,000 iters/sec on this machine.

It's definitely a notable improvement. I don't know why commenting out that single conditional makes such a big change. rpython's optimizations are amazing (and SLOW to perform) but also sometimes opaque in why little things can make such a big difference (at least they are to me).