[common] Removed tensor boundary checks in MXFP8 kernels (#1519)

Added constexpr checks of tensor boundaries
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

transformer_engine/common/util/cast_kernels.cuh

@@ -261,9 +261,15 @@ __global__ void __launch_bounds__(MXFP8_THREADS_PER_CHUNK)
               }
             }
             in_compute[j] = elt;
+
+            if constexpr (IS_ACT || IS_DACT) {
               if (!out_of_bounds) {
                 thread_amax = fmaxf(thread_amax, fabsf(elt));
               }
+            } else {
+              // If no activation, elt is 0 so we can safely do this
+              thread_amax = fmaxf(thread_amax, fabsf(elt));
+            }
           }
 
           __builtin_assume(block_amax >= 0);
@@ -320,9 +326,14 @@ __global__ void __launch_bounds__(MXFP8_THREADS_PER_CHUNK)
             }
           }
           in_compute[i] = elt;
+          if constexpr (IS_ACT || IS_DACT) {
             if (!out_of_bounds) {
               amax = fmaxf(amax, fabsf(elt));
             }
+          } else {
+            // If no activation, elt is 0 so we can safely do this
+            amax = fmaxf(amax, fabsf(elt));
+          }
         }
 
         __builtin_assume(block_amax >= 0);