add for loop-type constructor

Lots of people want for loop functionality for models, but they are very slow to define and execute. There are more efficient representations for each of these (e.g. vectorisation/linear algebra or adding an iteration-type method for dynamical models), but not all. It would be useful to have functionality for for loop -like code that uses Tensorflow's while loop syntax. Something like:

```r
fun <- function (i, arglist) {
arglist$x <- i ^ abs(arglist$x)
arglist
}

a <- normal(0, 1)
out <- loop(i = 1:10, fun, x = a)
```
I.e. the interface could be quite similar to lapply, but would behave cumulatively, updating arguments in the arglist. It would be equivalent to the R for loop:
```r
out <- list <- list(x = a)
for (i in 1:10) {
list <- fun(i, list)
out <- mapply(c, out, list, SIMPLIFY = FALSE)
}
```
I.e. returning a list, with vectors for the replicated elements (each with length 10 in this case).

This could use the same internal function creation trick as the ODE gist to create a tensorflow function, then map it to `tf.while_loop()` internally. Rather than spooling out the for loop into a really complex greta DAG/TF graph, this would iterate over the same piece of graph on the tensorflow side.