flat_draw
=========

.. py:function:: t3toolbox.backend.optimizers.flat_draw(problem, batch)

   .. code-block:: python

      def flat_draw(
              problem: Problem,
              batch:   int,        # measurements per minibatch
      ) -> typ.Callable:           # draw(rng) -> (sample_B, data_B)


   The **default** minibatch draw: a uniform random subset of ``batch`` measurements across the whole
   (flattened) sample stack ``W`` -- the robust default. Returns a ``draw(rng) -> (sample_B, data_B)``
   over ``problem``'s full data, via the kind's ``take``. A user may pass **any** ``draw`` instead (slice
   X / P / order, importance-sample, ...). Host numpy by default; pass jax data + a jax draw to keep the
   minibatch on device (the optimizer never compiles the draw -- only the per-step kernel).