Tract

* More memory/cache friendly execution order
* Several fixes around symbolic dimensions computation (some should help with attention models)

* [AMX] Put AMX for iOS behind a feature gate ("tract-linalg/apple-amx-ios").

* [ONNX] Support for external storage of tensors with offset and length
* [ONNX] Lots of fixes around binary quantized operators (add, mul, etc)
* [PY] Fix python source distribution
* [AMX] Activate AMX on iOS
* [API] Introduce transforms in external api
* [BLAS] Introduce a simple BLAS transform for Matrix multiplication
* [F16] Introduce a Reduce<MeanOfSquares> that solves many L2 normalization errors in f16

This version has been yanked to revert systematic activation of AMX on iOS. AMX is a private API and Apple may reject an App that performs AMX instructions.

* [ONNX] Support for external storage of tensors with offset and length

* MSRV is now 1.75.0
* [internal] ConvUnary and MatmulUnary are replaced by binary, potentially dynamic equivalent

* [ONNX] LayerNormalization support