Add optional graph-modification report to optimize()

[take-cheeze]

The optimization passes already compute how many positive transforms each pass applies to the graph (CountBasedPassAnalysis), but every layer discarded this information, so optimize() only returned the optimized model.

This threads a per-pass modification report through the whole stack:

• PassManagerAnalysis now carries a transform_counts map (pass name -> total positive transforms); both GeneralPassManager::run and FixedPointPassManager::run populate it instead of returning an empty analysis.
• Optimizer::optimize / Optimize / OptimizeFixed accept an optional report out-parameter (defaults to nullptr, fully backward compatible).
• New nanobind bindings optimize_report / optimize_fixedpoint_report and the path-based optimize_from_path_report / optimize_fixedpoint_from_path_report return the report alongside the model.
• Python optimize() gains a return_report keyword. When True it returns a (model, report) tuple where report is a dict mapping pass name to the number of graph modifications; the default return type is unchanged.

Add tests covering the tuple return, the counted transform, and the zero-count (pass ran but matched nothing) case.

Claude-Session: https://claude.ai/code/session_01TSYXLexqD5r4Lkg3KfeGZj

[take-cheeze]

For maintainers reviewing this: I wanted to add a concrete data point for why the per-pass report is worth having, beyond just observability.

A caller doing repeated/fixed-point optimization has to detect "did the last round change anything" somehow. The naive way is to keep the previous ModelProto around and either serialize+hash it or run MessageDifferencer::Equals against the new one. That's a full-message comparison every round, and it's not cheap: a model's serialized size is dominated by initializer weight data, which most passes (shape inference, most of onnx-optimizer's passes) don't touch on a given round, but a naive comparison still has to read all of it just to prove nothing changed.

We ran into exactly this in onnxsim, where this same patch is merged as onnxsim/optimizer#3 (identical diff to this PR). It's consumed in onnxsim/onnxsim#635: the innermost, most-frequently-run fixed point (OptAndShape = alternating shape inference and this optimizer) used to hash the entire serialized model every round to detect convergence -- expensive on models with large initializers, since it has to read all that weight data even on rounds that only touched a handful of nodes. With transform_counts, that check becomes "did any pass report a nonzero count" instead -- exact, and O(pass count) instead of O(model size). onnxsim/onnxsim#637 made this the default path and has real measured numbers: on mixer_l16_224_in21k (~300 MB of initializers), end-to-end simplification went from 268.7s -> 111.0s on the default path and 342.6s -> 85.1s on the graph-native path, validated against a 90-model regression sweep with 0 regressions.

So this isn't a hypothetical win -- it's load-bearing in a real downstream fixed-point optimizer today.
(I've used claude code to write this)