[Bug] Dynamic shape arrays and reshaping

It turns out that there are several bugs involved in attempting to reshape dynamic shaped arrays. In particular:

• jnp.reshape of a dynamically-shaped array with new compile-time shape results in the array being returned without any change in shape:

  @qjit(abstracted_axes={0: 'm', 1: 'n'}, keep_intermediate=True)
  def g(a):
       return jnp.reshape(a, (3, 5))
  
  a = jnp.ones([1, 3], dtype=float)
  g(a)
  

  I would expect the output to be of shape (3, 5) but with junk values for the out of bound elements.

• jnp.reshape of a dynamically-shaped array with new dynamic shape results in a segfault during compilation. Lowering to MLIR seems to happen correctly, but we suspect that mhlo.dynamic_reshape doesn't have a lowering rule.

  @qjit(abstracted_axes={0: 'm', 1: 'n'}, keep_intermediate=True)
  def g(a):
       return jnp.reshape(a, (a.shape[1], a.shape[0]))
  
  a = jnp.ones([1, 3], dtype=float)
  g(a)
  

  • Separately, we should consider adding a proper exception to catch this, rather than the segfault killing the kernel.

• jnp.reshape of a dynamically-shaped array within a loop iteration leads to a cryptic error even before compilation:

  >>> @qjit(abstracted_axes={0: 'm', 1: 'n'})
  ... def g(x):
  ...     @catalyst.for_loop(0, 10, 1, experimental_preserve_dimensions=False)
  ...     def loop(_, a):
  ...         return jnp.reshape(a, (3, 1))
  ...     return loop(x)
  >>> a = jnp.ones([1, 3], dtype=float)
  >>> g(a)
  ValueError: Too few leaves for PyTreeDef; expected 1, got 0
  

Originally posted by @josh146 in https://github.com/PennyLaneAI/catalyst/pull/904#discussion_r1667077464

Start by reproducing the three jnp.reshape cases using qjit, dynamic abstracted_axes, and catalyst.for_loop. Trace the MLIR lowering and the suspected mhlo.dynamic_reshape path, then verify that compile-time and dynamic reshapes behave as specified, loop reshaping no longer raises the PyTree error, and invalid cases produce a proper exception instead of a segfault.