Preserve folded dynamic-shape metadata through constant folding

[stevenvar]

This is an experimental follow-up to symbolic content propagation that teaches TensorFlow constant folding to preserve dynamic-shape provenance on folded constants by walking backward through their ancestor graph.

Today, when a value is constant-folded into a Const, we can lose the information that it originally came from a dynamic shape computation. This branch adds a recursive ancestor walk in constant folding so that if a folded value ultimately descends from a Shape / _xla_shape_derived path tied to an _Arg with dynamic dim expressions, we keep that inferred shape metadata on the replacement Const.

In short:

• before: only directly shape-derived folded constants kept dynamic metadata
• after: folded constants can also inherit that metadata from dynamic shape ancestors further upstream

This branch was split out as a separate experiment because it broadens provenance recovery beyond the direct-source behavior used in the main contents-propagation work.