[BUG Fix]: restrict max 2d block load heigth according to spec (#1152)

![image](https://github.com/intel/intel-xpu-backend-for-triton/assets/68101902/8598d5e4-d70e-491a-a042-91f386b5b3b3)
For 2d load, max height is 32(also for VNNI load). For those loads that
have height greater than 32, they will return all zeros. This will lead
to incorrect results.

third_party/intel/lib/TritonIntelGPUTransforms/MatchTargetSize.cpp

@@ -475,7 +475,8 @@ MatchTargetSizePass::getSubOpSize(RankedTensorType type) const {
         (isa<ttgi::WarpEncodingAttr, ttg::DotOperandEncodingAttr>(layout)) ? 32
                                                                            : 0;
     subSize[1] = (shape[1] > colLimit) ? colLimit : shape[1];
-    int64_t max = maxLoadStoreSize * 4 / sizeInBytes / subSize[1];
+    // FIXME: From gfxspec, max 2d block load height is 32
+    int64_t max = 32;
     subSize[0] = std::min(max, shape[0]);
   } break;
   default: