Add a normalize parameter to dense_diff_pool

[Hu-Chuxuan]

I did this modification because:

While I was using this function, I found the link_loss being exceptionally small and thus almost not contributing anything, making the graph structure hard to preserve and all nodes resulting in a few clusters. This is because adj.numel() are generally very large. After I recovered the result by multiplying adj.numel() back, results became reasonable.

According to the original paper, the link_loss is not divided by the size of adjacency matrix
image

After communicating with people on slack, I think it makes more sense for us to allow users to have an option of not being normalized.

[codecov]

Codecov Report

Merging #4847 (05a3e4c) into master (97c50a0) will increase coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head 05a3e4c differs from pull request most recent head 4a0fdc5. Consider uploading reports for the commit 4a0fdc5 to get more accurate results

@@           Coverage Diff           @@
##           master    #4847   +/-   ##
=======================================
  Coverage   82.71%   82.71%           
=======================================
  Files         326      326           
  Lines       17520    17521    +1     
=======================================
+ Hits        14491    14492    +1     
  Misses       3029     3029           

Impacted Files	Coverage Δ
torch_geometric/nn/dense/diff_pool.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 97c50a0...4a0fdc5. Read the comment docs.

[wsad1]

looks good to me.