Tract

* [NNEF/OPL] introduce json resource loader
* extend Complex number support (under a feature flag)

* [nnef] new identifier syntax is now opt-in for serialization (both accepted at loading)
* alpha-level C interface. how and how to deploy it (where to put the .h, whether or not to build and ship dylibs)
* alpha-level python interface. deployed on pypi as "tract". At this stage, API is undocumented and may still change significantly.

* [BREAKING] TValue are now used in run() instead of the previous mix of Tensor and Arc<Tensor>
* internal API breaking changes: no more op_families, libcli split away. State is no longer Send (but can be "frozen" to a Send counterpart).
* Symbols can now be String instead of char. They are not shared globally anymore, but scoped in the Model instead.
* [pulse] S symbol is no longer magic. The time dimension symbol must be provided at pulsification time.
* [pulse] In most cases, we can now pulsify without an explicit pulse len (pulse len can be expression).
* [cli] deprecated "x" syntax for shape is removed
* [nnef/opl] new i"..." syntax for escaping identifiers: i"some arbitrary string". Allow serialization of any ONNX model with any kind of string as node names.
* [ONNX] Signal processing operators (DTF, STFT, MelWeightMatrix, BlackmanWindow, HammingWindow, HannWindow)
* [ONNX] bitwise operations
* [ONNX] Compatibility target raised to operator set 18

* [NNEF] Introduce a "resource" extension for loading values from a separate source (as a config file)
* Workaround for cpu detection failure on FreeBSD / arm64
* Various bug fixes

* [pulse] improve convolution (and others) pulsification to avoid some unecessary buffering delay
* [cli] support multiple streaming inputs and outputs
* [ONNX] more relaxed Clip operator rules

* prepare NNEF for further tract-opl extension (resource support)
* more generic matmul
* optimise some EinSum cases as matmul