rolling joins with roll=TRUE and value -Inf

Discovered due to #1006.

``` R
require(data.table) ## 1.9.5
DT = data.table(x=c(-Inf, 3), y=1:2, key="x")
#       x y
#1: -Inf 1
#2:    3 2

# wrong result.. should be 2,1
DT[J(2), roll=TRUE]
#    x  y
#1: 2 NA
```
