[PyTorch] Add annotation_str benchmark (pytorch#96496)

To be used to evaluate performance of following improvements. Baseline numbers:

https://gist.github.com/swolchok/c8bcb92be1dc6e67c4f7efad498becd5

Differential Revision: [D43919653](https://our.internmc.facebook.com/intern/diff/D43919653/)

**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D43919653/)!
Pull Request resolved: pytorch#96496
Approved by: https://github.com/Skylion007

benchmarks/serialization/nested_annotation_str.py

@@ -0,0 +1,23 @@
+import torch
+import torch.utils.benchmark as benchmark
+
+MEMO = {}
+def create_nested_dict_type(layers):
+    if layers == 0:
+        return torch._C.StringType.get()
+    if layers not in MEMO:
+        less_nested = create_nested_dict_type(layers - 1)
+        result = torch._C.DictType(torch._C.StringType.get(), torch._C.TupleType([less_nested, less_nested]))
+        MEMO[layers] = result
+    return MEMO[layers]
+
+
+nesting_levels = (1, 3, 5, 10)
+types = (reasonable, medium, big, huge) = [create_nested_dict_type(x) for x in nesting_levels]
+
+timers = [benchmark.Timer(stmt='x.annotation_str', globals={'x': nested_type}) for nested_type in types]
+
+for nesting_level, typ, timer in zip(nesting_levels, types, timers):
+    print("Nesting level:", nesting_level)
+    print("output:", typ.annotation_str[:70])
+    print(timer.blocked_autorange())