$IterationLimit: count the current chain of rewrites, not every lookup

[tombouctou]

Fixes #1993.

Since #1477, Symbol.evaluate adds 1 to evaluation.iteration_count each time it rewrites a symbol by an ownvalue. It never subtracts it again, and the count is only reset at the start of a top-level evaluation. So the count is the total number of symbol lookups in the whole evaluation, not the length of one chain of rewrites. Any long computation that just looks up a variable 4096 times is aborted:

In[1]:= x = {a}; Do[x, {5000}]; x
$IterationLimit::itlim: Iteration limit of 4096 exceeded.
Out[1]= $Aborted

WMA returns {a} here. Its $IterationLimit reference page describes the limit as "the maximum length of evaluation chain used in trying to evaluate any expression". In practice, loading a package of a few thousand definitions with Get fails partway through.

This PR gives the step back when the lookup finishes, using try/finally. The count is then the depth of the current chain x -> y -> ..., which is what $IterationLimit should bound. Endless chains are still stopped. The doctest Block[{$IterationLimit = 20}, yin := yang; yang := yin; yin] and the tests added in #1477 pass unchanged.

New tests in test/builtin/test_evaluation.py:

• 30 separate lookups under $IterationLimit = 20 give the value. This test fails on master.
• A chain of 30 rewrites passes under a limit of 40.
• The same chain is aborted under a limit of 20.

pytest and the doctests give the same results as on master (the only failures are environmental ones, XML/ZIP/JPEG FileFormat and the CLI tests).

🤖 Generated with Claude Code

[mmatera]

LGTM