Pymatching

Support added for PowerPC architecture (92 and 94)

The numpy python dependency is now pinned to major version 1 (`numpy==1.*`) since the recent v2 release is not yet supported.

Support added for using pymatching as a C++ dependency (see changes in 77)

Renamed `retworkx` to `rustworkx` (since the [retworkx](https://pypi.org/project/retworkx/) package name has been deprecated):
* Added methods `Matching.load_from_rustworkx` and `Matching.to_rustworkx`
* DeprecationWarning added to `Matching.load_from_retworkx` and `Matching.to_retworkx`
* `retworkx` dependency removed. `rustworkx` has *not* been added as a dependency, and instead an `ImportError` is raised if `Matching.load_from_rustworkx` or `Matching.to_rustworkx` are used without `rustworkx` installed.

PyMatching now requires `Python>=3.7` (rather than `Python>=3.6`)

Fixes an issue where multiple boundary nodes could sometimes lead to boundary edges being merged incorrectly (issue 81, fixed in 86)

Features and improvements

Added the following methods to `pymatching.Matching`:
* `pymatching.Matching.decode_batch`: decodes a batch of (optionally bit-packed) shots to decode, instead of just a single shot (as done by `pymatching.Matching.decode`). For smaller/easier problems, `Matching.decode_batch` is _much_ faster than iterating over the shots in python and calling `Matching.decode` for each one. E.g. When decoding d=11,rounds=11, p=0.1% surface code circuits, looping over `Matching.decode` in python is 1us per round vs 0.37us per round using `Matching.decode_batch`. For d=5,rounds=5, p=0.01%, looping Matching.decode in python is 0.8us per round vs 0.02us per round with Matching.decode_batch.
* `pymatching.Matching.decode_to_edges_array`: returns the solution as an array of edges (given as pairs of detector node indices).
* `pymatching.Matching.from_detector_error_model_file`
* `pymatching.Matching.from_stim_circuit`
* `pymatching.Matching.from_stim_circuit_file`

Adds `pymatching.cli`, which provides access to the pymatching command line tool via the Python API.

Adds the `pymatching` console script entry point to the setup.py. Makes the `pymatching` command line tool available when pip installing.

Bug fixes

Fixes a bug in the command line tool where b8-formatted stim dets files couldn't be read on windows machines (also causing issues with sinter, which uses the command line tool).

This patch release fixes an issue where placing a detection event on a node marked as a boundary node in the Python API would lead to an exception being raised (issue 41).

Flagship changes:
* The C++ extension has been completely rewritten, and includes a new implementation of the blossom algorithm
* PyMatching v2.0 is now around 100-1000x faster than PyMatching v0.7 for decoding surface codes with circuit-level noise
* The new implementation is also _exact_. It does not make the "local matching" approximation used in v0.7 and earlier,
which resulted in a reduction in accuracy for very large surface code circuits
* The runtime of the new version is approximately linear in the size of the graph, and can process between 1 and 10
million detection events per second

Features and improvements
* Decoding speeds improved 100-1000x
* `pymatching.Matching.load_from_check_matrix` is around 30x faster
* Added methods to `pymatching.Matching`:
* `pymatching.Matching.add_boundary_edge`
* `pymatching.Matching.decode_to_matched_dets_array`
* `pymatching.Matching.decode_to_matched_dets_dict`
* `pymatching.Matching.has_edge`
* `pymatching.Matching.get_edge_data`
* `pymatching.Matching.set_min_num_fault_ids`

* Added static methods to `pymatching.Matching`:
* `pymatching.Matching.from_detector_error_model`
* `pymatching.Matching.from_check_matrix` (static method equivalent to constructing a new `pymatching.Matching` and
then calling `pymatching.Matching.load_from_check_matrix`)

* `merge_strategy` argument added to `pymatching.Matching.add_edge`, `pymatching.Matching.add_boundary_edge` and
`pymatching.Matching.load_from_check_matrix`. When an attempt is made to
add an edge already present in the graph, the following strategies are available:
* `disallow`: raises a `ValueError` if a parallel edge is added
* `independent`: edge weights and error probabilities updated treating the parallel edges as independent error mechanisms
* `smallest-weight`: The edge with the smallest weight is kept
* `keep-original`: The existing edge is kept and the new edge is silently ignored
* `replace`: Only the edge being added is kept

* `faults_matrix` argument added to `pymatching.Matching.from_check_matrix`. `faults_matrix` is an array that can be used to set
the `fault_ids` attribute of edges in a graph loaded from a check matrix. For example, `faults_matrix` can be used to specify the
logical observables whose outcome should be predicted by the decoder.

Minor changes to the API
* `spacelike_weights` renamed to `weights` in `pymatching.Matching.load_from_check_matrix`. But `spacelike_weights`
still accepted for backward compatibility.
* `H` renamed to `graph` in the constructor for `pymatching.Matching`. `H` still accepted for backward compatibility.
* `precompute_shortest_paths` argument removed from constructor (won't break if supplied)
* Both a virtual boundary node (via `pymatching.Matching.add_boundary_edge`) and a regular boundary node are supported.
These are treated equivalently by the decoder, but are treated differently by `pymatching.Matching.edges`,
`pymatching.Matching.to_networkx` and `pymatching.Matching.to_retworkx`

Minor breaking changes to the API
* `pymatching.Matching.matching_graph` renamed to `pymatching.Matching._matching_graph` (since this property is only
intended to be used internally to access the C++ extension)
* `pymatching.Matching.decode` arguments after the first argument are now keyword-only, and `num_neighbours` has been
removed as an argument. Users' code won't break or raise a warning if they already used keyword arguments (and `num_neighbours` is
silently ignored when provided). If users provide `num_neighbours` or `return_weight` as a positional argument, a deprecation warning is raised.
* `pymatching.Matching.compute_all_pairs_shortest_paths` method removed

Bug fixes
* Fixed a bug where timelike error probabilities were set incorrectly if provided as a scalar