Add detector support to CUDA-Q

[bmhowe23]

Overview

This PR introduces support for defining detectors in CUDA-Q kernels, a feature for Quantum Error Correction (QEC), and adds functionality to export kernels to the Stim circuit format.

Details

1. Detectors Support

• New Operation: Introduced cudaq::detector (C++) and cudaq.detector (Python) to define detectors. Detectors compare measurement results to check for errors (parity checks).
• IR & Dialect: Added quake.detector operation to the Quake dialect and its lowering to QIR (__quantum__qis__detector).
• Compiler: Added an EraseDetectors pass to remove detector operations when targeting backends that do not support them (e.g., standard hardware execution, or if explicitly undesired).
• Runtime:
  • Updated ExecutionManager and simulators to support tracing and handling detector operations.
  • Implemented cudaq.detectors() utility to extract detector metadata (measurement indices) from a kernel.

2. Stim Integration

• cudaq.to_stim(kernel): Added a new Python utility that converts a CUDA-Q kernel into a Stim circuit representation.
  • Supports gates, measurements, noise models, and the new detector operations.
• Stim Simulator: Updated StimCircuitSimulator to support a "to_stim" execution mode for string generation.

3. Tracing Improvements

• Measurement Tracing: Updated BasicExecutionManager to explicitly append measurement instructions (mz) to the kernel trace even in tracer mode.
  • Side Effect: The cudaq.draw output now explicitly shows measurement operations, updating expectations in test_kernel_features.py and draw_test.cpp.

Tests

• New Tests:
  • python/tests/backends/test_to_stim.py: Validates Stim circuit generation, including noise models and detectors.
  • python/tests/kernel/test_detectors.py: Verifies cudaq.detectors() analysis.
  • unittests/integration/detector_tester.cpp: C++ integration tests for detector IR and runtime behavior.
• Updated Tests: Adjusted existing visualization tests to account for the improved visibility of measurements in kernel traces.

[bmhowe23]

This is almost certainly too fragile or dangerous to keep it this way. If there is a standard MLIR way to do this, we should do it here.

[1tnguyen]

Just want to check that we'd have some info after ':' in the error message.

[schweitzpgi]

First, nice work pulling together this prototype from end to end. Very cool.

After talking to the architects, I think we should hold off merging this for now until we have a consensus on the user facing aspects, etc.