0.6.4

New Features

• Add ArgSort symbol/node which returns a stable argsort of its
  predecessor.
• Add expression() decorator for constructing expressions.
• Add the AccumulateZip symbol and corresponding AccumulateZipNode
  node in the C++. This symbol computes an arbitrary accumulate
  operation on 1-d arrays, similar to numpy.ufunc.accumulate, except
  that it takes any number of operands/predecessors. The accumulate
  operation is supplied by the user and encoded as an Expression.
• The Constant symbol is now able to pass ownership of the
  corresponding numpy array to the ConstantNode which prevents the
  numpy array from being garbage collected even after the symbol has
  been collected. This means it is no longer necessary to save
  references to all the numpy arrays used by Constant symbols on the
  model itself. To support this, a new constructor for the C++
  ConstantNode has been added which takes a special
  ConstantNode::DataSource struct allowing the Python reference
  count to modified when appropriate.
• Support automatic constant promotion for arithmetic operations on
  Symbols. This enables operations like model.binary() + 5, where
  the 5 will automatically be converted to a Constant symbol and
  added to the model.
• Add std::initializer_list and std::span overloads to several
  InputNode methods.
• Allow the user to provide an initial value to ArraySymbol.max(),
  .min(), .prod(), and .sum() methods.
• Add Resize symbol and C++ ResizeNode.
• Change the behavior of Model.constant() to keep a cache of
  symbols. Multiple calls to Model.constant() with the same array
  will no longer create redundant symbols.
• Add C++ deduplicate_diff_view range view.
• Add C++ type_list utility class.
• Added Mean symbol/node which returns the mean of its predecessor.
• Add support for Python 3.14.

Upgrade Notes

• Move C++ deduplicate_diff() function from
  dwave-optimization/utils.hpp to dwave-optimization/array.hpp.
• Change the signatures of the C++ ravel_multi_index() and
  unravel_index() functions to accept shapes rather than strides.
  This is consistent with NumPy.
• Remove C++ Array::View::at() method.

Bug Fixes

• Implement the C++ UnaryOpNode::sizeinfo() method to handle unary
  operations on dynamic arrays.
• There was an issue on some models with "diamonds" in the multi-DAG
  structure where adding subsequent nodes could trigger a exponential
  recursion of calling Array::min()/max(). This has been fixed by
  using a min/max cache by default in these methods.