perf: parallelize remapping to improve FTS compaction

[BubbleCal]

Tested it on MS MARCO and before this the remap can't be finished in 35m (I didn't wait for finishing because that's too long), and now it's 3.15min

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 37 lines in your changes missing coverage. Please review.

Project coverage is 78.04%. Comparing base (d2f636c) to head (2d08d8e).
Report is 4 commits behind head on main.

Files with missing lines	Patch %	Lines
rust/lance-index/src/scalar/inverted/builder.rs	0.00%	37 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2834      +/-   ##
==========================================
+ Coverage   77.99%   78.04%   +0.05%     
==========================================
  Files         229      229              
  Lines       70192    70314     +122     
  Branches    70192    70314     +122     
==========================================
+ Hits        54744    54878     +134     
+ Misses      12376    12338      -38     
- Partials     3072     3098      +26     

Flag	Coverage Δ
unittests	78.04% <0.00%> (+0.05%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[westonpace]

This seems like a good improvement. Did you setup any kind of benchmark that led you to the conclusion this needed to be sped up? It might be nice to have that benchmark checked in.

[westonpace]

I'm not sure there is any particular benefit to using for_each_init (versus for_each and declaring the new variables at the beginning of the closure) here since you are consuming the Vecs at the end of the closure but there is no harm in it either.

[BubbleCal]

Yeah it doesn't benefit from this, did this because I was thinking I would clear the vectors then I can reuse them but finally took it...