Fast single row predict API v2

[AlbertoEAF]

• Fix bug introduced in PR Feat/optimize single prediction #2992 for Fast predict (closes [bug] Introduced bug in *Fast() methods with the locking strategy PR #3241)
• Uses the newer API to make *Fast predict calls faster
  The latest single-thread benchmark with a model of 42 numerical features achieved a speedup 1.30x-1.85x vs the standard single row prediction API)
• Stricter API with const parameters

[AlbertoEAF]

@imatiach-msft @StrikerRUS @guolinke please review, it is now even faster by the latest tests.

It could be slightly faster if we used a fixed address for the input prediction data instance, but @imatiach-msft needs a dynamic one, so I left the code like this as the extra speedup was not that large.

Thank you for waiting for the patch and good luck with the v3 release! ;)

[StrikerRUS]

I'm confused by const/non-const params mismatch in C API functions.

[guolinke]

For the non pointer arguments (and non reference$), I think we can unify them to the non constant types.

[AlbertoEAF]

@StrikerRUS good catch! I had const mismatches even between the .c and .h in the function signatures which made things extremely weird.

@guolinke I put instead everything const just because many were lacking const when they could be. Now I think it's really easy to interpret from the user's viewpoint. Only the input booster - which will be modified - and output pointers, are non-const.

[StrikerRUS]

@AlbertoEAF Thanks a lot! I think with all const it looks much better.

[AlbertoEAF]

@guolinke do you think it's ready to merge?

Thank you

[guolinke]

Yeah, looks good to me

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.