Regression in SPEC2017 541.leela_r on AArch64

Since https://github.com/llvm/llvm-project/pull/204795 made "-use-dereferenceable-at-point-semantics=true" the default, we now fail to vectorize some code in the function `FastState::play_random_move()` of SPEC2017 541.leela_r.

This leads to a regression of ~2.5% measured on a Neoverse V2 machine.

Reproducer to follow.

Start by reading PR 204795 and inspect the vectorization behavior of SPEC2017 541.leela_r in FastState::play_random_move() on AArch64. Once the promised reproducer is available, run it and compare behavior with the dereferenceable-at-point semantics change; done means the code vectorizes again without the reported Neoverse V2 regression.

Written by the indexing model from the issue text.