As far as I am aware, cortical columns are still integral to Numenta’s HTM and TBT models. Although in recent years, it seems as though they are trying to pack more and more capability into the full column (as opposed to just L4). Among these additional capabilities is some form of grid-cell based rectification process (that might also involve some feedback through the thalamus) that would permit a single column to recognize a given feature regardless of its orientation. The exact interaction between the various layers is still not well understood, but several theories have been put forward.
The K-winner-take-all algorithm is a model meant to roughly approximate the localized inhibition effect of the basket and chandelier cells. The effect that they intended to model has been the enforcement of activation sparsity. K-winner-take-all is a cheap way to enforce sparsity, however, when in its most naive (and efficient) form, it does neglect some topological effects that the inhibitory cells provide (e.g. Calvin tiles). The HTM algorithm explicitly keeps the sparsity at the desired level rather than letting it arise naturally due to the balance of excitatory and inhibitory neurons.