[llvm] [AMDGPU] Use isGFX125xOnly as the assembler predicate for tensor load/store (PR #227887)

[Changpeng Fang via llvm-commits]

changpeng wrote:

Landed as 15d801334e531eb49f1624add593483d4fe9e020.

The merge went through, but GitHub returned a 502 before recording it on the PR, so this was left open. Since it was squashed, the head SHA (6b7baf6) never appears on main and GitHub's auto-detection won't mark it merged. Closing manually.

https://github.com/llvm/llvm-project/pull/227887