Add RLE dump code

[Kubuxu]

No description provided.

[codecov]

Codecov Report

Merging #7691 (4d8be81) into master (ab55a6f) will decrease coverage by 0.15%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #7691      +/-   ##
==========================================
- Coverage   39.56%   39.41%   -0.16%     
==========================================
  Files         637      637              
  Lines       67924    67979      +55     
==========================================
- Hits        26877    26796      -81     
- Misses      36445    36556     +111     
- Partials     4602     4627      +25     

Impacted Files	Coverage Δ
cmd/lotus-shed/sectors.go	0.00% <0.00%> (ø)
node/modules/dtypes/mpool.go	87.50% <0.00%> (-12.50%)	⬇️
chain/events/message_cache.go	87.50% <0.00%> (-12.50%)	⬇️
blockstore/api.go	24.00% <0.00%> (-8.00%)	⬇️
blockstore/blockstore.go	62.96% <0.00%> (-7.41%)	⬇️
chain/events/observer.go	71.64% <0.00%> (-6.72%)	⬇️
miner/miner.go	52.31% <0.00%> (-4.64%)	⬇️
chain/stmgr/execute.go	86.95% <0.00%> (-4.35%)	⬇️
markets/storageadapter/ondealsectorcommitted.go	77.33% <0.00%> (-4.00%)	⬇️
chain/stmgr/call.go	71.51% <0.00%> (-3.64%)	⬇️
... and 17 more

---

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 ab55a6f...4d8be81. Read the comment docs.

[arajasek]

Curious what motivated this

[Kubuxu]

I was analyzing how good/bad our RLE+ encoding is in practice.
Rough results:
A symbol in our case is a run.
Shannon Entropy of RLEs in AllocatedSectors: Η = 3.468 bits/symbol
RLE+ encoding: ΔH = 0.7687 bits/symbol
RLE+ with small modification: ΔH = 0.203 bits/symbol
RLE+ with Hufman: ΔH = ~0.0178 bits/symbol
RLE+ with Asymetric Numeral Systems: ΔH ~= 0.001 bits/symbol

We could save 18% of RLE+ storage but it is such a small fraction that it isn't worth the complexity right now (after collecting data I know that RLEs are only ~10MB of chain state, but the churn is frequent).
The major reason for the high ΔH of our RLE+ coding are 6-bit encodings of symbols 2 and 3.

If I were to design RLE+ today I would have gone with RLE with Huffman coding of small runs and additional data lengths.