Pyqrack

The gate `mcr()` ("multiply-controlled rotation") did not properly cast its angle parameter to `c_double`, when calling the shared C API; this has been fixed. (Binaries remain the same as v0.12.0.)

3 issues are addressed in this release:
1. `in_ket()` and `out_ket()` methods have been added to `QrackSimulator()`, for direct state vector writing and reading. We categorically discourage using these methods wherever avoidable, but automatic reactive Schmidt decomposition might still recover some separability from directly set state vectors.
2. Regression in `QPager` has been fixed, which previously prevented any paged simulations at all, above single OpenCL max allocation segment limits.
3. Per [the vm6502q/qrack README](https://github.com/vm6502q/qrack#qpager-distributed-simulation-options), `QRACK_QPAGER_DEVICES_HOST_POINTER` can specify a choice between OpenCL device memory or host memory, on a per-page basis.

With the use of `dump()` and `dump_ids()`, users can now directly output raw state vector amplitude values. (We discourage any recourse to state vector output unless one truly cannot operate without an explicit set of state vector amplitudes, or unless one would otherwise be compelled to reconstruct a state vector from measurement distributions.)

`dump_ids()` outputs the labels associated with state vector "wires" in permutation order. `dump()` outputs a list of complex number amplitudes, with corresponding "wire" order. These methods are not generally safe to use in threaded user code, (not including any internal Qrack threading). Also note that Qrack state vector amplitudes will typically have a randomized global phase offset, which is a matter of Qrack internal conventions and has no theoretically observable physical effect.

QEngine::ShuffleBuffers()` now relies on (semi-)async dispatch instead of `std::future` based threading. (This has been an aspirational goal, for several recent releases.)

SIMD based complex arithmetic has also been touched up, for performance.

Qrack has added internal decompositions for `mcmuln()` and `mcdivn()`, ("multiply-controlled multiplication/division modulo 'N'"). This might allow for better Schmidt decomposition.