Draft: ENH: Improve performance of np.broadcast_arrays (v2)

[eendebakpt]

This PR uses np.nditer to implement np.broadcast_arrays. The performance is good (also see #26160):

main:
broadcast_arrays(*args) 1.20 us
broadcast_arrays(*args_two) 7.26 us
broadcast_arrays(*args_multi) 9.11 us

PR #26160
broadcast_arrays(*args) 1.03 us
broadcast_arrays(*args_two) 3.83 us
broadcast_arrays(*args_multi) 3.87 us

PR:
broadcast_arrays(*args) 0.89 us
broadcast_arrays(*args_two) 0.99 us
broadcast_arrays(*args_multi) 1.15 us

A change in behaviour is that all output of np.broadcast_arrays is now readonly. This might be acceptable, as writeable broadcasted arrays have been deprecated for some time: #12609