Fwrite integer rownames

[ColeMiller1]

Closes #4957 .

The biggest change is behavior of quote = auto for default row names and for integer assigned row-names because this PR makes all row names characters.

library(data.table)
DT = data.table(foo=1:3,bar=c(1.2,9.8,-6.0))

# 1.14.0
fwrite(DT,row.names=TRUE, quote = 'auto')
## "",foo,bar
## "1",1,1.2
## "2",2,9.8
## "3",3,-6

#1.14.1 with PR. Double quote are lost
fwrite(DT,row.names=TRUE, quote = 'auto')
## "",foo,bar
## 1,1,1.2
## 2,2,9.8
## 3,3,-6

I believe this better matches existing behavior with quote = 'auto' as the current behavior will not add double quotes to character row.names:

row.names(DT) = letters[1:3]
fwrite(DT,row.names=TRUE, quote = 'auto')

## "",foo,bar
## a,1,1.2
## b,2,9.8
## c,3,-6

Finally, based on this PR, this part of fwrite.c shown below is no longer used. In the issue, I comment about how to do this more directly in C. I would be happy to move forward with a C approach but it just seems like a higher diff for little productivity. While maybe big data people who need performance use row names, overall I doubt it.

data.table/src/fwrite.c

Lines 876 to 881 in 831013a

	if (args.rowNames==NULL) {
	if (doQuote!=0/*NA'auto' or true*/) *ch++='"';
	int64_t rn = i+1;
	writeInt64(&rn, 0, &ch);
	if (doQuote!=0) *ch++='"';
	} else {

[mattdowle]

I like your reasoning for this; i.e. option 1 you detailed in issue. However I'm just a bit concerned about any inadvertent usage that ends up somehow calling row.names=TRUE without the user intending to. Then 1:nrow will be coerced to character by this line and the global character cache gets clobbered. Would prefer not to leave that door open. Another door open would be benchmarkers who either deliberately or by accident display results with row.names = TRUE and show poor performance.
Hence I went for option 3 as you described.

[ColeMiller1]

Brilliant! Thanks for your time - changes look great!