Keep symbolic arithmetic inference scoped to dynamic XLA

[stevenvar]

Summary

TensorFlow-side dynamic value inference has a symbolic path for shape-like values produced by Add, Sub, Mul, and Div. That path previously ran even when dynamic value inference was disabled, and it returned early when symbolic propagation could not handle the available operands. This could prevent GraphProperties from continuing through its normal value-inference logic.

This change:

• runs symbolic arithmetic propagation only when dynamic value inference is enabled;
• returns early only after symbolic propagation succeeds;
• otherwise falls through to TensorFlow's existing inference path.

Ordinary GraphProperties behavior is therefore preserved when dynamic value inference is disabled, while dynamic XLA inference can still propagate compatible symbolic shape values.

Tests

• //tensorflow/core/grappler/costs:graph_properties_test
• all 63 tests pass

[Guillermo-Callaghan]

This still does not fall back when symbolic arithmetic fails 😒. Shape-value tensors can validly produce -1, for example, [1] - [2], but InferenceContext::Subtract interprets these values as dimensions and returns InvalidArgument. Because TF_RETURN_IF_ERROR propagates that error, GraphProperties fails instead of reaching normal value inference, contrary to the new fallback behaviour. Please treat arithmetic failure as “symbolic propagation does not apply” (or use tensor-value semantics) and add a regression test covering negative results and unknown-shape sentinels.

[stevenvar]

Thank you for your review. As you correctly said, -1 is invalid as a concrete dimension but remains a perfectly valid integer tensor value. The symbolic path now steps aside instead of failing GraphProperties, and regression tests ensure this distinction should not require another reminder.

[Guillermo-Callaghan]

LGTM