incorrect bias correction if bias.correct.control$split is not sorted

[brianstock-NOAA]

Sorry for not taking the time to give a reproducible example... but hopefully this is described well enough for now.

If I specify bias.correct.control$split using Which as in TMBhelper::fit_tmb (lines 130-148), I get incorrect answers if Which is not sorted.

1. Bias correct all (default):

# mod is optimized TMB model
sdreps <- list()
sdreps[[1]] = summary(TMB::sdreport(mod, bias.correct=TRUE))

2. Bias correct only a subset of derived quantities, one at a time (very slow, ~4x the above run time)

to.correct <- c("log_SSB", "log_SSB_FXSPR", "log_FAA_tot", "log_FXSPR", "log_pred_catch")
Which = as.vector(unlist(mod$env$ADreportIndex()[ to.correct ] ))
sdreps[[2]] = summary(TMB::sdreport(mod, bias.correct=TRUE, bias.correct.control=list(sd=FALSE, split=Which)))

3. Bias correct same subset, but hoping to speed it up by using one chunk, split = list(Which)

sdreps[[3]] = summary(TMB::sdreport(mod, bias.correct=TRUE, bias.correct.control=list(sd=FALSE, split=list(Which))))

4. Bias correct same subset, but sort the indices, split = list(sort(Which))

sdreps[[4]] = summary(TMB::sdreport(mod, bias.correct=TRUE, bias.correct.control=list(sd=FALSE, split=list(sort(Which)))))

Look at bias corrected SSB:

ssb.bc <- lapply(sdreps, function(x) {
	ind.ssb <- which(rownames(x) == "log_SSB")
	bc <- exp(x[ind.ssb,3])
	return(bc)
})
ssb.bc

I find that 1, 2, and 4 give the same bias corrected SSB, but 3 is different (wrong). 2 is very slow, but robust to issues where there are a few NAs (won't return all NAs if there is one NaN). 4 is fast and correct, and should probably be the default in TMBhelper::fit_tmb if BS.control[["vars_to_correct"]] is used. Or maybe TMB::sdreport should check for this?