[PrefetchBlock] Do not use location from an erased loop (#3345)

Closes #3314

The creation of barriers in this pass uses a location obtained from the
`loop` entity that [was erased
earlier](https://github.com/intel/intel-xpu-backend-for-triton/blob/cc1774086e9ec75567b3d0b6ab4d54de3e4fba6e/third_party/intel/lib/TritonIntelGPUTransforms/PrefetchBlock.cpp#L404).
Surprisingly, it works on Linux but fails on Windows. Replaced the
erased `loop` with `newLoop`.

Windows CI with `TritonIntelGPU/prefetch-block.mlir` test passing:
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/13138880536/job/36660853060#step:13:789

Signed-off-by: Dmitry Chigarev <[email protected]>

Author: dchigarev

third_party/intel/lib/TritonIntelGPUTransforms/PrefetchBlock.cpp

@@ -454,7 +454,7 @@ void PrefetchBlockPass::injectPrefetchOpsInBody(
 
   // FIXME: try to use a named barrier to increase performance.
   if (injectSplitBarriers) {
-    Location loc = loop.getLoc();
+    Location loc = newLoop.getLoc();
     b.setInsertionPoint(yield);
     b.create<spirv::INTELControlBarrierWaitOp>(loc, spirv::Scope::Workgroup,
                                                spirv::Scope::Workgroup,