The Guard's eval Rereads Multiply the Loop Scan's Quadratic Cost #2440

Problem

The guard's _loop_parts scan is quadratic in the number of loop keywords on develop. The eval reading (#2080) re-reads each eval payload, which multiplies that cost on a command whose quoting the guard cannot follow. The eval read budget caps it, and none of the self-test timing rows uses many loop keywords.

Reproduce (constructed)

The Python expression below builds the command. The trailing line makes its quoting unknown:

'eval "a\'b"; ' * 5 + 'while true; ' * N + 'sleep 1\necho "$(echo "it\'s")"'

Measured by the review: at 24 KB develop takes 1.17s and the branch 4.36s, and at 48 KB develop takes 4.6s and the branch 10.8s. Both allow it.

Done looks like

The loop scan is linear, or the eval reading stops re-scanning loops it already read, and a self-test timing row with many loop keywords stays under the suite's limit.

Found by a local strict review pass on the #2080 branch.