Add na="null_specials" to differentiate Inf/NaN from NA

[r2evans]

I'm being optimistic that you'll be amenable to addressing #461. It's an opt-in addition to na= that defaults to unchanged legacy behavior.

Previous behavior:

toJSON(c(3.14, NA, NaN, 21, Inf, -Inf))
# [3.14,"NA","NaN",21,"Inf","-Inf"] 
toJSON(c(3.14, NA, NaN, 21, Inf, -Inf), na = "null")
# [3.14,null,null,21,null,null] 
toJSON(c(3.14, NA, NaN, 21, Inf, -Inf), na = "string")
# [3.14,"NA","NaN",21,"Inf","-Inf"] 

With this patch:

toJSON(c(3.14, NA, NaN, 21, Inf, -Inf))
# [3.14,"NA","NaN",21,"Inf","-Inf"] 
toJSON(c(3.14, NA, NaN, 21, Inf, -Inf), na = "null")
# [3.14,null,null,21,null,null] 
toJSON(c(3.14, NA, NaN, 21, Inf, -Inf), na = "string")
# [3.14,"NA","NaN",21,"Inf","-Inf"] 
toJSON(c(3.14, NA, NaN, 21, Inf, -Inf), na = "null_specials")
# [3.14,null,"NaN",21,"Inf","-Inf"] 

I added unit-tests to confirm this behavior. All other tests pass.

In full disclosure, there is a slight performance dip. In the C code I tried to be a bit more "unrolled" to mitigate this, but it's non-zero.

set.seed(42)
n <- 10000
dat <- data.frame(lgl=sample(c(F,T,NA), size=n, replace=TRUE), int=sample(100L, size=n, replace=TRUE), num=runif(n=n), dat=sample(c(Sys.Date(), NA), size=n, replace=TRUE), psx=sample(c(Sys.time(), NA), size=n, replace=TRUE))
microbenchmark::microbenchmark(toJSON(dat), times=1000) # before this PR, after, and with "null_specials"
# Unit: milliseconds
#                             expr      min       lq     mean   median       uq      max neval
#               toJSON(dat) before 10.01105 10.26181 10.80194 10.32343 10.39608 78.23144  1000
#                toJSON(dat) after 11.40054 12.04596 12.74151 12.18012 12.41033 37.34489  1000
#  toJSON(dat, na="null_specials") 11.85302 12.49873 13.24990 12.62183 12.91810 39.65528  1000

The same method using n <- 10, a more modest comparison, not demonstrating a significant change.

n <- 10
### ...
# Unit: microseconds
#                             expr     min       lq     mean   median      uq       max neval
#               toJSON(dat) before 183.762 197.8250 215.3376 206.3735 220.539  2900.955  1000
#                toJSON(dat) after 182.819 189.4815 213.9801 193.9505 215.209  9884.157  1000
#  toJSON(dat, na="null_specials") 176.710 194.6475 217.5185 198.7885 216.521 10203.055  1000