hash_info: don't use dataclass

[efiop]

We are creating lots of these instances and dataclass is significantly slower.

For example, running:

from dvc_data.hashfile.hash_info import HashInfo


hashes = []
for value in range(10000000):
    hashes.append(HashInfo("md5", value))

before this PR takes ~6.1sec and after ~4.9sec.

Most of the improvement is actually coming from using __slots__, which were finally supported by dataclasses in 3.10.

Another (temporary) downside to using dataclass is that it is kinda buggy with other tools. E.g. mypyc is failing to compile it throwing an error like mypyc/mypyc#921 for the HashInfo.name.

[skshetry]

We are creating lots of these instances and dataclass is significantly slower.

How much perf improvements do you see with this?

[codecov-commenter]

Codecov Report

Merging #56 (a90646e) into main (070fd08) will increase coverage by 0.03%.
The diff coverage is 76.47%.

@@            Coverage Diff             @@
##             main      #56      +/-   ##
==========================================
+ Coverage   34.24%   34.28%   +0.03%     
==========================================
  Files          33       33              
  Lines        1673     1677       +4     
  Branches      261      262       +1     
==========================================
+ Hits          573      575       +2     
- Misses       1089     1090       +1     
- Partials       11       12       +1     

Impacted Files	Coverage Δ
src/dvc_data/hashfile/hash_info.py	56.41% <76.47%> (-0.74%)	⬇️

---

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 070fd08...a90646e. Read the comment docs.

[efiop]

@skshetry Great question! Added to the PR desc.

[efiop]

Overall we are not heavy users of dataclasses in this application (and a few more places like TreeEntry), so the overhead for a simpler implementation is very minor.

[skshetry]

I just wish we had not done this kind of change without looking into other alternatives (eg: attrs with slots, dataclasses, etc). This makes debugging quite hard. Also, the above benchmarks seem unfair.