avoid quadratic per-leaf subtree walk in is_complex_subscript

[sahvx655-wq]

Profiling a file that packs a long run of implicitly concatenated strings into a single subscript or list (x = ["s0" "s1" ... "sn"]) showed the whole run going quadratic in the default style, with Line.is_complex_subscript and pre_order dominating the trace. append calls it once per leaf while a square bracket is open, and when the bracketed body is one node that holds no comparison or arithmetic sub-node the any(... in subscript_start.pre_order()) walk never short-circuits, so each of the n string leaves re-walks the entire subtree. A comma-separated subscript dodges this because subscriptlist narrows to the relevant child through child_towards, but a single implicitly concatenated string sits directly under the bracket and gets walked in full every time. It is reachable pre-auth through blackd, and left unfixed a 4000-string list takes several seconds where a comma list of the same size is near-instant.

The walk result depends only on subscript_start, which does not change while a line is being built, so I cache it per node on the Line. Keeping the cache at this layer rather than having callers skip the call preserves the exact spacing behaviour is_complex_subscript drives for genuine complex slices, and the subscriptlist branch still narrows per child so each distinct child is walked once. Output is byte-identical across the test data and the black source tree in stable and preview modes, and the same 4000-string input drops from ~3.4s to ~0.18s with the curve going from quadratic to linear.

[github-actions]

diff-shades results comparing this PR (9cc6741) to main (51abf53):

--preview style: no changes

--stable style: no changes

---

What is this? | Workflow run | diff-shades documentation