v0.3.5

Added

• Gate fusion for the QCSim and GPU backends, controlled by gate_fusion
  in SimulatorConfig and exposed through the simulator, network, C and
  native request interfaces with fusion statistics.
  SimulatorConfig.gate_fusion is optional: None means unset, and
  GetConfiguration("gate_fusion") then reports "auto".
• Native MPO queries on the QCSim and GPU backends: bit-vector probabilities,
  density matrix elements, dense density-matrix output, complex Pauli
  expectations and local operators A rho A†, through IState, the C
  interface, the Python bindings and the native request API.
• trim() and recanonicalize() for matrix-product states, and the trim
  and recanonicalize maintenance actions for MPS in the request API.
• Swap optimization (look-ahead routing) for MPOs, and
  disable_optimized_swapping and lookahead_depth for MPO requests.

Changed

• Breaking: the Python methods trim_mpo and recanonicalize_mpo are
  renamed trim and recanonicalize; they now work on MPS as well as MPO.
• When gate_fusion is not set, QCSim statevectors below 11 qubits and density
  matrices below 5 qubits run unfused, as fusion is slower there. An explicit
  true or false still forces fusion on or off at any size.
• MPO max_bond_dimension accepts zero for unlimited, and the trace-restore
  and hermitize options now apply to GPU MPO too.
• precision and use_double_precision are aliases kept in sync. Qiskit Aer
  honours single precision for initialization and state export, and the
  network no longer forces single precision.
• Faster CPU fused circuits, from the vectorized dense gate kernels and a
  cheaper fusion cache.
• The GPU diagnostics operation accepts an MPS for maintenance. An MPS reports
  no diagnostics by default.

Fixed

• Circuit layering no longer moves a measurement ahead of an earlier write to
  the same classical bit, or ahead of a conditional gate still reading it.
  Routed MPS/MPO runs returned wrong distributions for circuits that write a
  bit twice.
• Single-shot network calls keep the shot's classical bits when the executed
  circuit is narrower than the network allocation; ExecuteOnHost reported all
  zeros.
• MPS look-ahead routing survives gate fusion, so fused MPS circuits with
  mid-circuit measurements no longer fall back to local routing.
• FlushSimulator waits for queued GPU statevector and density-matrix work, and
  fusion boundaries no longer stall the GPU.
• Multithreading flags are forwarded to QCSim MPS and MPO simulators.
• Tensor network clones are isolated from the original.
• GPU, distributed GPU, Aer and composite simulators ignore a precision set
  after initialization instead of throwing. A new precision takes effect after
  Clear and the next Initialize.
• CI build installs apt packages after refreshing the package index.

Dependencies

• QCSim pin bumped to 1edb647: vectorized dense gate kernels, MPS
  recanonicalization fixes, MPO long-chain scaling, MPS sampling and input
  validation fixes, and Windows compile fixes.