Update base for Update on "Add ops (addr) | feat(torchlib)"

-----

- Mark nan==nan in assert_close
- Fix aten_isfinite by allowing the exporter to do type promotion according to #563 and introduce `TFloatHighPrecision` for it
- Fix typo in `TRealOrUInt8`

[ghstack-poisoned]

Author: justinchuby

onnxscript/tests/function_libs/torch_lib/ops_test_data.py

@@ -431,14 +431,7 @@ def _where_input_wrangler(
     TorchLibOpInfo("abs", core_ops.aten_abs_complex, complex=True),
     TorchLibOpInfo("acos", core_ops.aten_acos),
     TorchLibOpInfo("acosh", core_ops.aten_acosh),
-    TorchLibOpInfo(
-        "add",
-        core_ops.aten_add,
-    ).xfail(
-        dtypes=[torch.float16],
-        reason="fixme: float16 failed, tensor-likes are not close for FullGraph mode. https://github.com/microsoft/onnxruntime/issues/15977",
-        test_class_name="TestOutputConsistencyFullGraph",
-    ),
+    TorchLibOpInfo("add", core_ops.aten_add, tolerance={torch.float16: (1e-3, 1e-3)}),
     TorchLibOpInfo("addbmm", core_ops.aten_addbmm),
     TorchLibOpInfo("addcdiv", core_ops.aten_addcdiv),
     TorchLibOpInfo("addcmul", core_ops.aten_addcmul),