Closes #4455
hi @jangorecki tagging you for a review since this is a fix for #4455 which you labeled.
Turns out the fix was really easy, just had to delete two lines of code.
I had to update the expected values for two existing tests, which I suspect were created based on the empirical output of melt, rather than the theoretically expected values.
In the linked issue the example code which illustrates the problem is:

library(data.table)
wide.dt <- data.table(
  family=1,
  age_child1=NA_real_, sex_child1=NA_character_,
  age_child2=5, sex_child2="m")
measure.vars <- list(age=c(2,4), sex=c(3,5))
(na.rm.T <- melt(wide.dt, measure.vars=measure.vars, na.rm=TRUE))
(na.rm.F <- melt(wide.dt, measure.vars=measure.vars, na.rm=FALSE)[!is.na(age)])

on current master we get results that are inconsistent (na.rm=TRUE gives variable=1, but na.rm=FALSE gives variable=2).

> (na.rm.T <- melt(wide.dt, measure.vars=measure.vars, na.rm=TRUE))

   family variable age sex
1:      1        1   5   m
> (na.rm.F <- melt(wide.dt, measure.vars=measure.vars, na.rm=FALSE)[!is.na(age)])
   family variable age sex
1:      1        2   5   m

This is problematic for downstream applications such as tdhock/nc#11 which need variable to be the corresponding index of the measure.vars list elements. In the example above we need variable=2 in both cases so that we know the data come from child2.

After applying this PR I get

> (na.rm.T <- melt(wide.dt, measure.vars=measure.vars, na.rm=TRUE))
   family variable age sex
1:      1        2   5   m
> (na.rm.F <- melt(wide.dt, measure.vars=measure.vars, na.rm=FALSE)[!is.na(age)])
   family variable age sex
1:      1        2   5   m

All tests pass on my computer (R-4.0.2, Ubuntu 18.04), please code review.