Rebased onto upstream triton main with the merged warp-pipeline PR #10840
(always-on sched.barrier after each stage's memory ops,
SchedGroupMask::non_mem_non_sideeffect). Drops the tutorial's local
fence_loads/keep_order commits (superseded by #10840); keeps the out-of-tree
LLIR-scheduler / amdgcnas plugin infra. LLVM pin: 62b7cf96 -> 850a2b1b.