Pulled out from discussions in #8886:

• Preallocate the result array (MAINT: Block algorithm with a single copy per call to block #11971)
• Alias np.b_[...] to np.block([...])- shorthands are useful for nesting calls
• Implement block destructuring, np.b_[...] = somearr (comes cheap after preallocation)
• Add a np.block(arr : ndarray[object]) overload, that takes an ndarray of blocks that align in all axes (ie, such that the concatenation order does not matter)
  • add a grid=True parameter to the same effect (ENH: Require that arguments to block lie on a grid eric-wieser/numpy#2)
• Allow broadcasting
  • Extra leading dimensions: With two stacks of arrays, A.shape == (3, 1, 2, 2), B.shape = (1, 4, 2, 2), should np.block([A, B]).shape == (3,4,2,4)? (currently: ValueError)
  • All dimensions
• Deprecate bmat in favor of np.block(...).view(np.matrix)