join performance regression

introduced between https://github.com/h2oai/datatable/compare/a45cc503494571bfbf0feb00ece03ad0bab16dfc...31fbad7b471113e30f34de8cc2b321c76f926c32

reproduced using
```sh
cd db-benchmark
source ./pydatatable/py-pydatatable/bin/activate
pip install --upgrade git+https://github.com/h2oai/datatable.git@a45cc503494571bfbf0feb00ece03ad0bab16dfc
./_launcher/solution.R --solution=pydatatable --task=join --nrow=1e7 --quiet=true --out=dt-time.csv
sleep 5
./_launcher/solution.R --solution=pydatatable --task=join --nrow=1e8 --quiet=true --out=dt-time.csv
sleep 5
pip install --upgrade git+https://github.com/h2oai/datatable.git@31fbad7b471113e30f34de8cc2b321c76f926c32
./_launcher/solution.R --solution=pydatatable --task=join --nrow=1e7 --quiet=true --out=dt-time.csv
sleep 5
./_launcher/solution.R --solution=pydatatable --task=join --nrow=1e8 --quiet=true --out=dt-time.csv
deactivate
R
```
```r
library(data.table)
fctr = function(x) factor(x, levels=unique(x)) ## this retains order for strings
d = fread("dt-time.csv")
dd = d[run==1L, .(time_sec, in_rows, question=fctr(question), git=fctr(substr(git,1,7)))]
dcast(dd, in_rows+question~git, value.var="time_sec")
```
```
in_rows question a45cc50 31fbad7
1: 10000000 small inner on int 7.419 10.852
2: 10000000 medium inner on int 12.999 13.902
3: 10000000 medium outer on int 8.128 6.900
4: 10000000 medium inner on factor 14.145 10.085
5: 10000000 big inner on int 16.316 27.621
6: 100000000 small inner on int 39.371 154.445
7: 100000000 medium inner on int 33.088 208.082
8: 100000000 medium outer on int 24.478 94.655
9: 100000000 medium inner on factor 33.831 163.471
10: 100000000 big inner on int 114.547 153.472
```