Update on "group constraints by arg"

Before, we would emit a soup of specializations / constraints without any obvious order to guide readability.

With this diff, we group such results by arg, and add comments preceding each group. Empirically, the results read much better.

Differential Revision: [D45995199](https://our.internmc.facebook.com/intern/diff/D45995199/)

cc voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx desertfire

[ghstack-poisoned]

torch/fx/experimental/symbolic_shapes.py

@@ -1778,7 +1778,6 @@ def remove_default_lower_bound(dc):
         for i, arg in enumerate(signature.parameters.keys()):
             args_index[arg] = (i, arg)
 
-        buf = ""
         def print_results(grouped, indent, result_fn):
             nonlocal buf
 
@@ -1792,6 +1791,7 @@ def print_results(grouped, indent, result_fn):
                 for result in results:
                     buf += f"\n{indent}{result_fn(result)}"
 
+        buf = ""
         indent = 4 * " "
         if self._static_results:
             grouped_static_results = group(self._static_results, args_index)
@@ -1811,7 +1811,7 @@ def print_results(grouped, indent, result_fn):
             buf += f"\n{indent}return ["
             print_results(
                 grouped_dynamic_results,
-                indent*2,
+                indent * 2,
                 lambda result: f"{remove_default_lower_bound(result)},",
             )
             buf += f"\n{indent}]\n```\n"