gmedian retain class

[mattdowle]

Closes #3079
Follow up to #3546
gmedian now retains the class, other than for integer64 since it always returns double.

Also reduced lines of code in gmedian by adding an integer64 case to quickselect.c, moving the 2nd-half minimum loops into quickselect.c, and finding location of median inside quickselect rather than passing that in. Existing test 1579 covered int, double, integer64 and na.rm cases nicely.

R API raised outside loops in gmedian (since altrep R 3.5 greater overhead).

A few trailing spaces removed (my editor does that on save).

A few unrelated tests tweaked to pass in dev mode when data.table is not installed.

Checked all gforce function retain attributes. Added comment to the only one that doesn't (gvarsd1) since class unlikely applicable for sd/var.

[codecov]

Codecov Report

Merging #3564 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3564      +/-   ##
==========================================
+ Coverage   97.45%   97.46%   +<.01%     
==========================================
  Files          66       66              
  Lines       12806    12678     -128     
==========================================
- Hits        12480    12356     -124     
+ Misses        326      322       -4

Impacted Files	Coverage Δ
R/fwrite.R	98.75% <100%> (ø)	⬆️
src/gsumm.c	96.9% <100%> (+0.16%)	⬆️
src/quickselect.c	100% <100%> (ø)	⬆️

---

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 93f50f7...2533c90. Read the comment docs.

[mattdowle]

I declared isInt64 as const bool so that branch prediction should be very good. Interesting to test but, thanks to branch prediction, I doubt there will be a measurable difference in this case.
It is probably worth creating totally branch free cases; i.e. taking isunsorted and irowslen outside too. But that would complicate/expand the code. The move of the R API outside the loops should have already had a greater benefit.
However, I knew that when gather() is applied this serial gather will go away. Then gmedian can apply the quickselect in parallel. That's what should be done next.