v4.21.0

Features

• TrajectorySimulator, a Monte-Carlo trajectory simulator for programs with mid-circuit
  measurements and resets. It samples pure-state trajectories under JAX, returning the
  measurement outcomes alongside the final state; sample streams trajectories in batches,
  data-parallel across the available devices, so the shot count is not bounded by device
  memory. compute is jax.jit-traceable (though not differentiable).
  (#1861)
• A gate angle may be any Quil arithmetic expression over declared memory, so the parametric
  programs quilc emits -- RX(theta[0]/2 + pi) 0 -- simulate without rewriting.
  (#1869). Restored from
  #1860, where it was errantly dropped during a
  rebase (prior to the 4.19.0 release).
• The simulators accept an initial_state to evolve from, so a circuit can be split in two and
  a prefix shared across many runs evolved once instead of once per run.

Fixed

• Tolerant Circuit equality (and its unhashability, which follows), the split of
  Circuit.__post_init__ into named coercion and validation steps, and a mypy fix for JAX's
  dynamically defined config flags. Restored from
  #1860, where it was errantly dropped during a
  rebase (prior to the 4.19.0 release).
• Circuit.to_kraus_maps no longer truncates the Kraus set.