Updated docstring for mesh initialization

 * not allocated at once, but by parts (by dimensions). The flow is roughly the

 * following:

 *

 *  - Allocate vertices and set their physical coordinates, deallocate input

 *    array of points.

 *  - Allocate cells and set their subvertex indices (but not other subentity

 *    indices), deallocate cell seeds (the cells will be used later).

 *  - Initialize vertices by copying their physical coordinates from the input

 *    array, deallocate the input array of points.

 *  - Initialize cells by copying their subvertex indices from cell seeds (but

 *    other subentity indices are left uninitialized), deallocate cell seeds

 *    (the cells will be used later).

 *  - For all dimensions D from (cell dimension - 1) to 1:

 *     - Create intermediate entity seeds, count the number of entities with

 *       current dimension.

 *     - Allocate entities and set their subvertex indices. Create an indexed

 *       set of entity seeds and reference orientations (if applicable).

 *     - Set their subvertex indices. Create an indexed set of entity seeds and

 *       reference orientations (if applicable).

 *     - For all superdimensions S > D:

 *        - Iterate over entities with dimension S and initialize their

 *          subentity indices with dimension D. Inverse mapping (D->S) is

 *        - For entities with dimension D, initialize their superentity indices

 *          with dimension S.

 *     - Deallocate all intermediate data structures.

 *  - For all superdimensions S > 0, repeat the same steps as above.

 *

 * Optimization notes:

 *   - Recomputing the seed key involves sorting all subvertex indices, but the