Create performance model for reduction operators

[gabeweisz]

Fixes #533

The performance model is currently bare-bones, based on O(n) operations being strictly necessary for these reductions.

It could be more tuned but that would depend significantly on the implementation and the theoretical limit is something like n/2 + log(n) at a minimum which is not likely significantly more accurate

[Copilot]

Pull request overview

Adds a first-pass performance model and categorization for single-GPU reduction-style ATen operators, and updates regression artifacts to reflect the new modeling in generated perf reports.

Changes:

• Introduce Reduce / aten_reduce perf model classes to estimate FLOPs/bytes for reduction-like ops.
• Map common aten:: reduce ops to the new perf model and categorize them under a new "Reduce" category.
• Add a new checked-in MI300 perf report reference .xlsx used by perf-report regression tests.

Reviewed changes

Copilot reviewed 2 out of 9 changed files in this pull request and generated 3 comments.

File	Description
TraceLens/PerfModel/perf_model.py	Adds Reduce and aten_reduce perf model implementation (FLOPs/bytes + param parsing).
TraceLens/PerfModel/torch_op_mapping.py	Registers reduce ops → aten_reduce, adds Reduce category and categorization branch.
tests/traces/mi300/Qwen_Qwen1.5-0.5B-Chat__1016005_perf_report.xlsx	New reference perf report for regression testing.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.