relax fp16 tolerance, forgot to include on "[ONNX][dynamo_export] Skip instance_norm decomp for export"

Otherwise, instance_norm is decomposed into batch_norm with training set to True.
Downstream exporter has no way to figure out that training is actually not needed.
On the other hand, ONNX does have InstanceNormalization operator defined, however
due to decomp, it unnecessarily exports as batch norm and glue code.

Depends on microsoft/onnxscript#1284

[ghstack-poisoned]

Author: BowenBao

test/onnx/test_fx_op_consistency.py

@@ -1957,6 +1957,7 @@ class TestOnnxModelOutputConsistency(onnx_test_common._TestONNXRuntime):
         "nn.functional.hardsigmoid": [1e-3, 5e-3],
         "nn.functional.hardswish": [1e-3, 5e-3],
         "nn.functional.hinge_embedding_loss": [4e-1, 3e-3],
+        "nn.functional.instance_norm": [1e-2, 1e-3],
         "nn.functional.interpolate": [1e-2, 1e-3],
         "nn.functional.kl_div": [2e-3, 2e-4],
         "nn.functional.multilabel_soft_margin_loss": [4e-2, 5e-3],