perf(ptv3): simplify SerializedPooling export clustering

[mojomex]

Summary

This changes the PTv3 SerializedPooling export path to derive pooling clusters directly from the existing serialized order instead of using the export-only unique path.

It also adds a focused module test that checks export mode against train mode for SerializedPooling.

What changed

• remove the export-path dependency on unique in SerializedPooling
• build cluster, indices, and idx_ptr by scanning pooled-code boundaries along the existing serialized order
• keep the existing pooled-order/inverse behavior unchanged after clustering
• add projects/PTv3/tests/test_serialized_pooling.py to verify export-mode and train-time outputs match for the changed module

Validation

• pytest -q projects/PTv3/tests/test_serialized_pooling.py

Benchmark context

This exact SerializedPooling change was benchmarked separately on an inference box. That rerun measured GPU total replicate-mean latency at 24.868 ms versus the original 26.937 ms baseline, a 2.069 ms (7.68%) improvement.

The benchmark/profile artifacts are intentionally not part of this PR.