[Frontend] Make the C++ trace the sole main TOG path; drop legacy ONNX TOG

[YWHyuk]

Summary

Make the C++ trace (trace.so + trace_cycles.tsv) the sole main TOG path and drop
the legacy ONNX Tile-Operation-Graph from the frontend codegen and the
frontend<->TOGSim connection.

• extension_codecache: emit only trace.so/trace_cycles.tsv (build-skip keys on
  trace.so); remove the legacy ONNX generation block and the TORCHSIM_LEGACY_TOG
  opt-in.
• TOGSimulator.run_standalone: always drive TOGSim with --trace_so. The ONNX
  --models_list branch is KEPT solely for the STONNE sparse path (extension_op.py).
• origins (which FX nodes a kernel came from) is preserved: logged per kernel run
  and recorded as a trailing "# origins:" line in trace_cycles.tsv (the C++
  cycle-table loader stops at the comment, so the current parser is unaffected).
• Drop the dead tog_file param; migrate scripts/chiplet.sh to --trace_so.
• Refresh CLAUDE.md TOG-generation notes.

Out of scope (kept)

• STONNE sparse path (tog_generator/onnx_utility + ONNX) is untouched.
• TOGSim C++ (TileGraphParser, onnx/protobuf) is untouched -- separate PR.

Verification

• test_add / test_matmul pass via the trace path (no tile_graph.onnx emitted).
• ILS mode (TORCHSIM_TLS_MODE=0, all iterations) passes.
• chiplet (c2_chiplet) runs on the trace path (Total execution cycles produced).

Known limitation

The trace path stubs per-tensor addresses in TOGSim (build_trace_tilegraph), so
chiplet NoC / DRAM-partition accuracy is approximate until the trace path consumes
real addresses (TOGSim-side follow-up). --attributes_list was already not a valid
Simulator option.