fix(ipld): fix reconstruction by re-enabling re-importing of rsmt2d.ExtendedDataSquare

[Wondertan]

An update of rsmt2d happened in #737 during Mamaki launch preparation. This update changed our code not to reimport the square on each reconstruction attempt which silently sneaked in the problem, which is not reproducible with current tests. Reconstruction tests(#702) reproduce it, but they are not yet merged.

I was able to change the code to reproduce it reliably. Mainly I had to fetch only from columns and force the code to download only [2:4) quadrants. Unfortunately, it is hard to extract the reproducible case in a separate test, so we can either rely on reconstruction tests that are soon to be merged, or I can still try to make a test.

The reason for the bug is explained in #787, as well as a cleaner way to fix it.
Closes #787

P.S. Was pretty hard to debug this case. I went on the wrong paths like 3 times. Though collected some minor findings that I am going to submit further.

[liamsi]

I would really, really prefer a testcase for this but I trust your hours of debugging this.

[Wondertan]

Force push with a cleaner commit

[codecov-commenter]

Codecov Report

Merging #804 (0dc7d7a) into main (a422ce1) will increase coverage by 0.03%.
The diff coverage is 72.72%.

@@            Coverage Diff             @@
##             main     #804      +/-   ##
==========================================
+ Coverage   53.15%   53.19%   +0.03%     
==========================================
  Files         117      117              
  Lines        6750     6760      +10     
==========================================
+ Hits         3588     3596       +8     
- Misses       2798     2801       +3     
+ Partials      364      363       -1     

Impacted Files	Coverage Δ
ipld/retriever.go	87.76% <72.72%> (-2.93%)	⬇️
header/core/listener.go	52.83% <0.00%> (-5.67%)	⬇️
header/p2p/server.go	56.52% <0.00%> (+3.26%)	⬆️
fraud/bad_encoding.go	34.44% <0.00%> (+3.33%)	⬆️

---

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 a422ce1...0dc7d7a. Read the comment docs.

[Wondertan]

@liamsi, I added a test anyway. You can try removing the fix commit and reproduce it yourself. Sometimes it fails with other rsmt2d errors besides byzantine. The test ensures that at least 2 more quadrants are fetching besides the initial one, and all of those written into the shared square, and somewhere in between, the reconstruction fails.

[Wondertan]

As always on CI something goes not like locally, aaaaa

[Wondertan]

Force push with an updated test that should not fail on CI, lets see...

[Wondertan]

Ok, the test is fixed on CI