docs(3-pool): correct CI-fn gradient scaling write-up

Match the n_ci fix and correct the imp-min explanation in the distributed
walkthrough's §10: the autograd-aware all_reduce backward SUM-reduces the grad
(it does not "broadcast unchanged"), and PPGD's CI grad carries an extra n_ci to
survive the CI-pool AVG (its V/U grad does not).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: danbraunai-goodfire