Pyqrack

This releases is concurrent with Qrack v7.1.1, which fixes a bug in `QUnit` multi-qubit buffer commutation.

This release fixes an edge case in the Schmidt decomposition "layer" when operating "over" stabilizer layer, which sometimes caused incorrect results. (Specifically, when Schmidt decomposition was previously applied to stabilizer subsystems, it was possible that a two-qubit subsystem would be separated while only updating the probability amplitude caches of one of the two qubits, whereas both need to be updated.)

Work on v0.18.0 prompted attention to the basic implementation of `QStabilizerHybrid` in the underlying C++ Qrack library. This uncovered a conceptual mistake in how we previously attempted measurement through non-Clifford single qubit gate buffers. Now, the minimum necessary edge cases of non-Clifford measurement resort to state vector representation before measurement. This improves accuracy (hopefully to the point of exact universal `QStabilizerHybrid` simulation, up to machine epsilon) at the expense of performance.

Improvements in stabilizer simulation have been made with the "reverse t-injection gadget" in Appendix A of [Fast Estimation of Outcome Probabilities for Quantum Circuits](https://journals.aps.org/prxquantum/abstract/10.1103/PRXQuantum.3.020361). These method improvements can be turned on or off for a `QrackSimulator` with the method `set_t_injection(bool)`. (They are on by default.) While implementing this, other minor improvements were made in `QStabilizerHybrid`.

Deadlocks in highly Schmidt-decomposed benchmarks seem to have been due to [failure of C++ Qrack to adhere to OpenCL standards for callbacks.](https://stackoverflow.com/questions/43831183/opencl-callback-hangs-freezes-deadlock-pthread-cond-wait#answer-45173284) Per the standard, recent work with OpenCL error handling and retry have been amended to also avoid any and all blocking calls in OpenCL callbacks. Also according to the standard, kernel dispatch on callbacks is explicitly flushed. (Retrospectively, v0.17.6 offers no practical improvement over v0.17.5.)

Recent performance improvements uncovered a deadlock condition (which was longstanding, but previously rarely encountered) in the asynchronous CPU dispatch. This has been fixed and tested in the underlying C++ library.