Add Qwen3 MoE route-trace generation and replay

[Happymic]

What This PR Does

This PR connects the generic replay harness to Qwen3 MoE routing. It adds the scripts needed to build Qwen route traces from real model inputs/weights and replay those traces through the emulator for pilot timing runs.

Concrete Changes

• Adds Qwen route-trace generation from tokenized benchmark inputs and local model weights.
• Adds a Qwen trace replay test that builds deterministic emulator workloads from those traces.
• Adds a batch runner for many traces with resume, skip-existing, keep-going, and failure logging.
• Adds result export utilities for pilot timing summaries.
• Adds artifact pruning so successful runs do not keep large .pt/memory dump files unnecessarily.
• Uses the repo-derived REPO_ROOT for subprocess cwd instead of a hardcoded local path.

What This PR Does Not Do

• It does not commit model weights.
• It does not commit generated route traces or timing outputs.
• It does not claim final GPQA/BFCL results. It adds the pilot harness needed to produce them.

Why This Is Separate

P1 (#94) proves replay on controlled traces. This PR makes replay useful for the real Qwen MoE workload path.

Validation

Ran locally:

export PLENA_ROOT=/path/to/plena
source "$PLENA_ROOT/venvs/plena-py311/bin/activate"
export PYTHONPATH="$PWD:$PWD/PLENA_Compiler:$PWD/PLENA_Tools:$PWD/transactional_emulator/testbench${PYTHONPATH:+:$PYTHONPATH}"
python -m compileall -q transactional_emulator/testbench/emulator_runner.py transactional_emulator/testbench/window1_p2
python -m transactional_emulator.testbench.window1_p2.build_route_traces --help
python -m transactional_emulator.testbench.window1_p2.run_trace_batch --help
python -m transactional_emulator.testbench.window1_p2.qwen3_trace_replay_test --help
python -m transactional_emulator.testbench.window1_p2.generate_true_routing_with_weights --help

After the portable cwd cleanup:

python -m compileall -q transactional_emulator/testbench/window1_p2
python -m transactional_emulator.testbench.window1_p2.run_trace_batch --help