Accelerate DGL csr neighbor sampling (#13588)

* Speedup and fix bug in dgl_csr_sampling op

* Update dgl_graph.cc

* simplify functions.

* avoid adding nodes in the last level in the queue.

* remove a hashtable lookup in neigh_pos.

* reduce a hashtable lookup in sub_ver_mp.

* merge copying vids and layers.

* reduce hashtable lookup when writing to output csr.

* fix a bug.

* limit the number of sampled vertices.

* fix lint.

* fix a compile error.

* fix compile error.

* fix compile.

* remove one hashtable lookup per vertex and hashtable iteration.

* remove queue.

* use vector for neigh_pos.

* fix lint

* avoid init output arrays.

* fix tests.

* fix tests.

* update docs.

* retrigger

* retrigger