feat: add shrinked target in linknet + dilation in postprocessing

[charlesmindee]

Similarly to the DB, this PR includes a shrink of the polygons in the target and a dilation in the postprocessing of the linknet, to prevent boxes to be merged on the heatmap because their boundaries are too close.
After some experiments, this lead to much better results.
Any feedback is welcome!

[codecov]

Codecov Report

Merging #822 (6ce142c) into main (612e4f8) will decrease coverage by 0.12%.
The diff coverage is 76.59%.

@@            Coverage Diff             @@
##             main     #822      +/-   ##
==========================================
- Coverage   96.10%   95.97%   -0.13%     
==========================================
  Files         131      131              
  Lines        4949     4988      +39     
==========================================
+ Hits         4756     4787      +31     
- Misses        193      201       +8     

Flag	Coverage Δ
unittests	95.97% <76.59%> (-0.13%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
doctr/models/detection/linknet/base.py	88.79% <76.59%> (-4.72%)	⬇️

---

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 612e4f8...6ce142c. Read the comment docs.

[fg-mindee]

Perhaps we could do this in a smoother way by sharing the box shrink+dilation ops between the DB & LinkNet?

Because I feel like we're only adding duplicate lines of code here 😅

[charlesmindee]

I modified the typo, for the code duplicates there is only the polygon_to_box part in common, I suggest refactoring that in another PR

[fg-mindee]

Thanks Charles!