Qualcomm: cap inf replacement value to fix 16a16w accuracy regression

PR #19660 folded ReplaceInfValues into QnnQuantizer._replace_inf and made
the inf stand-in equal to the full quant range. For 16a16w that is 65535
(vs the previous fixed 255), which blows up the attention-mask quant scale
and breaks stories110M decoding in test-llama-runner-qnn-linux. Cap the
magnitude at 255 to restore prior behavior; 8a8w is unaffected.

Co-authored-by: Claude <noreply@anthropic.com>

Author: psiddh

backends/qualcomm/quantizer/quantizer.py

@@ -416,7 +416,12 @@ def _get_quant_range(self, node):
                 if quant_info.output_qspec.quant_min is None
                 else quant_info.output_qspec.quant_min
             )
-            return quant_range
+            # Cap the inf stand-in so it does not dominate the tensor's
+            # dynamic range.  For >8-bit activations the full range (e.g.
+            # 65535 for uint16) would blow up the attention-mask quant scale
+            # and wreck accuracy; 255 keeps a reasonable scale for
+            # Llama-style attention masks.
+            return min(quant_range, 255)
 
     def _get_candidates_with_infinity_args(self, graph_module: GraphModule):
         binary_op_sources = [