feat: Embarrassingly parallel TASO

[aborgna-q]

Use the CircuitChunks splitter for running TASO on different parts of the circuit.

• Adds a --split-circ flag to the taso optimizer bin, to split the circuit before running instead of sharing the full circuit between threads.
• Adds a node_cost parameter to the splitter to split circuits according to our cost functions.

I'm getting some really good results with this, we should test if they are correct :P
Optimizing barenco_tof_10.json (192 CX gates):

• Single thread, 60s timeout:
  • min cost: 166 CX gates
  • peak memory usage: 8.48 GB
• 2 threads (96-CX chunks), 60s timeout:
  • min cost: 141 CX gates
  • peak memory usage: 8.06 GB
• 2 threads (96-CX chunks), 30s timeout followed by single thread for 30s:
  • min cost: 126 CX gates
  • peak memory usage: 8.01 GB

(these are noisy results, as I'm running them on a non-idle laptop)

---

Edit: Updated measurements

Optimizing barenco_tof_10.json (192 CX gates):

• Single thread, 60s timeout:
  • min cost: 138 CX gates
  • peak memory usage: 3.20 GB
• 2 threads (96-CX chunks), 60s timeout:
  • min cost: 126 CX gates
  • peak memory usage: 3.06 GB
• 2 threads (96-CX chunks), 30s timeout followed by single thread for 30s:
  • min cost: 124 CX gates
  • peak memory usage: 3.05 GB

[lmondada]

Give me a second, I'm testing equality on barenco_tof_10.json. I'm getting that they are not equivalent but I think it's a false positive.

[lmondada]

Looking good! We just need to find that bug...

[aborgna-q]

Update: with #150 the tests now use ~2.5GB peak memory.

[lmondada]

I'm tempted to approve it, but let's wait until I have verified the failing circuits using other means.

We'll resolve this next week.

[lmondada]

Can you run this one more time on barenco_tof_10 end-to-end and confirm it works? Then we can merge this in.

Thanks a lot!

[aborgna-q]

This still doesn't support identity wires in the chunks, so it will errors out if an optimization produces one. I'm working on that on the side, but that can go in a separate PR.

The current impl works alright as long as we don't do that.

Here are the updated benchmarks, optimizing test_files/barenco_tof_10.json (192 CX gates):

• Single thread, 60s timeout:
  • min cost: 138 CX gates
  • peak memory usage: 3.20 GB
• 2 threads (96-CX chunks), 60s timeout:
  • min cost: 126 CX gates
  • peak memory usage: 3.06 GB
• 2 threads (96-CX chunks), 30s timeout followed by single thread for 30s:
  • min cost: 124 CX gates
  • peak memory usage: 3.05 GB

[lmondada]

🚀

[aborgna-q]

The empty-chunks panic is now fixed in #172