perf: Don't always take all data buffers when gathering views

[orlp]

Always blindly taking all data buffers of all input chunks leads to exponential behavior in the new join. Now if there is more than one chunk we deduplicate and only take those data buffers we actually reference.

[codspeed-hq]

CodSpeed Performance Report

Merging #20435 will not alter performance

_{Comparing orlp:dedup-gather (39ac18b) with main (f213c36)}

Summary

✅ 41 untouched benchmarks

[codecov]

Codecov Report

Attention: Patch coverage is 80.70175% with 33 lines in your changes missing coverage. Please review.

Project coverage is 79.01%. Comparing base (2ae01fb) to head (39ac18b).
Report is 38 commits behind head on main.

Files with missing lines	Patch %	Lines
...tes/polars-ops/src/chunked_array/gather/chunked.rs	80.70%	33 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #20435      +/-   ##
==========================================
+ Coverage   78.98%   79.01%   +0.02%     
==========================================
  Files        1562     1562              
  Lines      220103   220221     +118     
  Branches     2486     2488       +2     
==========================================
+ Hits       173855   173997     +142     
+ Misses      45675    45651      -24     
  Partials      573      573              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.