Revert "Add stash_type=1 to RMSNormalization for FP32 variance accumulation"

This reverts commit ea4e95c.

Author: justinchuby

src/mobius/components/_rms_norm.py

@@ -43,7 +43,6 @@ def forward(self, op: builder.OpBuilder, hidden_states: ir.Value):
             effective_weight,
             epsilon=self.variance_epsilon,
             axis=-1,
-            stash_type=1,
         )
 
 
@@ -190,7 +189,4 @@ def apply_rms_norm(op: builder.OpBuilder, x, weight, eps):
     Returns:
         Normalized tensor with the same shape as input.
     """
-    # stash_type=1 (FLOAT) ensures the variance computation uses FP32
-    # internally even when the input is FP16/BF16, preventing overflow
-    # when squaring large values.
-    return op.RMSNormalization(x, weight, epsilon=eps, axis=-1, stash_type=1)
+    return op.RMSNormalization(x, weight, epsilon=eps, axis=-1)