Remove uphill bias from random walk controller (50/50)

Change directional bias from 60/40 (favoring increase) to 50/50
(no directional bias). The A/B comparison showed 8 workers was 5.2%
slower than 4 on 200K heaps, and the uphill bias was actively guiding
the walker toward a worse operating point.

Per Alex's direction: "no bias."

Forward-ported from commit 95d98be on parallel_gc_backport (3.12).
3.15 routing: controller body is in Python/gc.c rather than
Modules/gcmodule.c.

Author: SonicField

Python/gc.c

@@ -2304,9 +2304,9 @@ gc_collect_region(PyThreadState *tstate,
 
                 // 20% chance to step ±1 (proactive exploration)
                 if (rand_val < 0.2) {
-                    // Bias uphill: 60% chance to increase, 40% to decrease
+                    // No directional bias: 50/50 chance to increase or decrease
                     double dir_val = (double)((rng >> 16) & 0xFFFF) / 65535.0;
-                    int delta = (dir_val < 0.6) ? 1 : -1;
+                    int delta = (dir_val < 0.5) ? 1 : -1;
 
                     // Always step when the dice fires. Good values stick
                     // because they don't trigger further corrective steps.