Clean up graph passes and use common IR passes

[justinchuby]

Use IR apis and common passes. Eliminate the use of getattr in the touched files.

Fixed a bug where CSE would cause the same value to duplicate as graph outputs, which is invalid in onnx.

Removal of legacy helpers and predicates:

• Removed unused or redundant helper predicates and functions such as _is_elem, _count_consumers, _find_next_consumer_idx, _nodes, and _inputs from both the test imports and test implementations, favoring direct use of IR containers and methods.

Direct use of IR API:

• Replaced custom value constructor V_ir with the standard ir.val throughout all tests, ensuring consistency with the IR API.
• Simplified graph and function construction by using ir.Function and passing functions directly to ir.Model, removing complex logic for function attachment.

Test logic and assertion modernization:

• Updated all node and input/output container accesses to use direct iteration (list(g) or g.inputs, g.outputs) instead of helper functions, and used direct attribute access for names.
• Updated assertions to reflect correct ONNX semantics, e.g., graph outputs must remain distinct objects, and adjusted tests accordingly.

Documentation update:

• Updated the IR optimizer authoring guide to recommend using built-in IR methods and containers directly, and to avoid unnecessary helpers.

@enpasos

[justinchuby]

The only difference is this: The CSE pass in onnx-ir does not consider non deterministic ops. We may want to include them as an update.

[enpasos]

@justinchuby Thank you for the massive effort on this modernization — it’s absolutely the right direction for the codebase.

Unfortunately, I wasn’t able to get this branch to pass 100% of the regression tests in my environment. The remaining failures were primarily related to aggressive DCE pruning in nested conditionals, as well as some complex shape-inference edge cases (notably FFT / Scatter).

To unblock the migration, I’ve opened a derivative PR #152 (use-passes-stable) that cherry-picks your work but selectively reverts a few optimizations (DCE and global replacements) back to their legacy, safer behavior. The intent is to preserve correctness first while keeping the successful pass-based modernization intact.