Revert "[DEBUG] Revert "Enable SPV_INTEL_fp_fast_math_mode (#4058)"(#4473)" (#4576)

This reverts commit 38a1984.

Known cases of impact on accuracy of the following models: detectron2
and doctr_reco_predictor from
#4412 on PVC
and LayoutLMForSequenceClassification from
#4509 on ARL

---------

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

third_party/intel/triton_xpu.cc

@@ -300,8 +300,8 @@ void init_triton_intel(py::module &&m) {
       for (Instruction &inst : instructions(func)) {
         if (auto *op = dyn_cast<FPMathOperator>(&inst)) {
           FastMathFlags FMF;
-          // Allow contract when default fp fusion is enabled.
-          if ((enableFpFusion.has_value() && enableFpFusion.value()) &&
+          // Default to allow contract when default fp fusion is not disabled.
+          if ((!enableFpFusion.has_value() || enableFpFusion.value()) &&
               !fastMath.has_value()) {
             if (op->getOpcode() == Instruction::FAdd ||
                 op->getOpcode() == Instruction::FMul)