inductor: sort the reads buf by name (pytorch#89744)

Sort `read_writes.reads` by name to make sure the same graph is generated for a fixed model. Otherwise, the buffer reuse may be different since the order of `read_writes.reads` is random.

Pull Request resolved: pytorch#89744
Approved by: https://github.com/jgong5, https://github.com/jansel

torch/_inductor/scheduler.py

@@ -318,7 +318,9 @@ def allocate(self):
             from .codegen.triton_template import should_use_template
             from .codegen.wrapper import buffer_reuse_key
 
-            for read in self.read_writes.reads:
+            ordered_reads = sorted(self.read_writes.reads, key=lambda x: x.name)
+
+            for read in ordered_reads:
                 input_node: BaseSchedulerNode = self.scheduler.name_to_node.get(
                     read.name
                 )