Add row names to null data.table on assign

[ColeMiller1]

Closes #4597.

I will do the news release after 1.13.0 is released. Also, will update test tomorrow to make this pass.

library(data.table)

dt = data.table()
dt[, new := 1:5]
attr(dt, "row.names") ##fixed!
#> [1] 1 2 3 4 5

## warning for now - easy to change to error with feedback. 
## data.table::groupingsets depended on this so there may be other issues with a hard change
dt = data.table(x = integer())
dt[, new := 1:5]

## Warning message:
## In `[.data.table`(dt, , `:=`(new, 1:5)) :
##   The data.table has 0 rows but there are 5 rows being assigned. The value will be ignored but new column(s) will be 
##   added. This warning will be upgraded to error in next release
dt
#> Empty data.table (0 rows and 2 cols): x,new

[codecov]

Codecov Report

Merging #4609 (7a5263b) into master (b284e37) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #4609   +/-   ##
=======================================
  Coverage   99.42%   99.42%           
=======================================
  Files          73       73           
  Lines       14450    14455    +5     
=======================================
+ Hits        14367    14372    +5     
  Misses         83       83           

Impacted Files	Coverage Δ
src/assign.c	99.70% <100.00%> (+<0.01%)	⬆️

---

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 b284e37...7a5263b. Read the comment docs.

[ColeMiller1]

@jangorecki let me know if I should split this up. The updating of row.names of data.table()[, new_var := 1:5] makes sense. However, I'm uncertain about data.table(x = integer())[, new_var := 1]. Current behavior is a zero-length variable being assigned. Proposed behavior is a warning with long-term aim of being an error. I did it as a single PR because they were relatively closely related.

[mattdowle]

Great fix, thanks.
As you alluded to above, and influenced by recent #4262, I went ahead and removed the new warning and reverted the changed tests. I'm (now) viewing data.table(x = integer())[, new_var := 1] as recycling of the length-1 to match the length-0 data. Happily, this is already what the tests tested, so my change to this PR brings that side of it back to current behavior so it can be merged to fix the row.names bug. Can always reconsider, discuss, and have another PR in future as you said.