Fix composite pass missing captured values in ops with nested regions

[mstojkovicTT]

Problem

The BuildStableHLOCompositePass performs a reverse graph traversal from output boundary markers to input boundary markers to collect all ops that belong to a composite. The traversal followed op->getOperands() to discover dependencies, but this only captures values passed directly to an op, and it misses values that are used inside nested regions (e.g., the body of a stablehlo.reduce) but defined in the parent function.

This caused the generated composite impl function to reference values that were never included as arguments or cloned ops, leading to invalid MLIR.

Fix

After processing each op's direct operands, the traversal now also inspects all nested regions using mlir::getUsedValuesDefinedAbove() (mlir doc) to find captured values. These are handled with the same logic as direct operands:

• Block arguments added as composite arguments
• Constants included in the impl function body
• Other ops added to the traversal worklist

[mstojkovicTT]

Tagging @qihqi and @ysiraichi for visibility

[ysiraichi]

Could you add some tests that show the failure? The build failure is detailed in this comment.
Again, since I'm not familiar with this module, I can't really review it.