fix NaN handling in parquet statistics

[crepererum]

Which issue does this PR close?

Closes #255.

Rationale for this change

Fixes NaN handling in parquet statistics. This is in line with the C++ stack:

https://github.com/apache/arrow/blob/b3e43987c47b2f01b204a2d954f882f7161616ef/cpp/src/parquet/statistics_test.cc#L1000-L1043

What changes are included in this PR?

Filter out NaN values from statistics + tests.

Are there any user-facing changes?

Yes: formally NaN were included in the stats but at "random" (i.e. when the data started with an NaN than the min/max values are NaN, otherwise min/max are non-NaN). Now the behavior is: NaN are excluded always. If the only NaN values are present, then min/max are unset.

[crepererum]

Should we introduce some generic T::is_nan function into the data type abstraction layer (that basically returns False for every non-float/double type and calls .is_nan() for float/double) or is this local workaround good enough?

[nevi-me]

I don't think that's necessary. It'll just be adding overhead for us.

The compiler can figure out what we're trying to do, and will optimise out the branch for anything that's not a floating point number.

Look at this example: https://godbolt.org/z/z7M3Mfqzz

The is_nan::<i32>() always returns false, and the ucomisd does the floating point comparison

[codecov-commenter]

Codecov Report

Merging #256 (acd4f52) into master (8f030db) will increase coverage by 0.01%.
The diff coverage is 90.00%.

@@            Coverage Diff             @@
##           master     #256      +/-   ##
==========================================
+ Coverage   82.52%   82.53%   +0.01%     
==========================================
  Files         162      162              
  Lines       43672    43786     +114     
==========================================
+ Hits        36039    36139     +100     
- Misses       7633     7647      +14     

Impacted Files	Coverage Δ
parquet/src/column/writer.rs	93.76% <90.00%> (-0.27%)	⬇️
parquet/src/arrow/levels.rs	81.25% <0.00%> (-0.44%)	⬇️
parquet/src/arrow/arrow_writer.rs	98.42% <0.00%> (+0.09%)	⬆️

---

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 8f030db...acd4f52. Read the comment docs.

[alamb]

I believe this fixes #255 rather than #225 so I updated the description accordingly

[alamb]

The change (and tests) look really good to me. @sunchao or @nevi-me do you have any thoughts?

[alamb]

Thanks @crepererum !

[nevi-me]

LGTM