[release/internal/3.5.x] Fix null-ptr crash in TritonAMDGPUOptimizeDotOperands when tt.trans takes load result directly

[mstankov-amd]

[AMD] Fix null-ptr crash in TritonAMDGPUOptimizeDotOperands when tt.trans takes load result directly

Problem

TritonAMDGPUOptimizeDotOperands crashes with PassManager::run failed when
compiling the flex attention backward kernel on gfx120x. The
failure is a null-pointer dereference in the ReuseShmemForDirectAndTransposedUse
pattern.

When tt.trans uses the load result directly,
transOp.getSrc().getDefiningOp<ConvertLayoutOp>() returns nullptr. The code
then passes this null cvtOp into rewriter.modifyOpInPlace(cvtOp, ...),
crashing the pass manager. There is also a latent null-dereference in the LDBG
line immediately after (debug builds only).

Fix

One-line null check: if no ConvertLayoutOp precedes tt.trans, skip this user
with continue. The pattern then falls through to the existing if (!transDot)
guard and returns notifyMatchFailure gracefully, leaving the IR unchanged.

cvtOp = transOp.getSrc().getDefiningOp<ttg::ConvertLayoutOp>();
if (!cvtOp)
  continue;

Ref #ROCM-1778

The file has changed in upstream/main, so this should be only 3.5 specific change. There is no failure on version 3.6