Use withTimeout when resolving remote future may corrupt its state

I need to execute some machine learning algorithms on a remote GPU server. It is likely that some heavy training is going and all CPU and GPU resources are occupied. In this case, starting a remote process and communicating with it are both likely to hang. Therefore I need an external timeout (rather than an remote call on `withTimeout` as suggested in #169) to control the time it takes and if timeout occurs other measures will be taken.

```r
library(future)

v <- R.utils::withTimeout({
p <- remote({
Sys.sleep(10)
1
}, workers = "", user = "", persistent = FALSE, earlySignal = TRUE)

value(p)
}, timeout = 1)
```

However, the behavior can be quite random as I run it repeatedly. In some cases, it works as expected, but sometimes it directly returns `1` without a sleep, and in other cases, it ends up in the following error:

```
Error: Unexpected result (of class ‘NULL’ != ‘FutureResult’) retrieved for ClusterFuture future (label = ‘’, expression = ‘{; Sys.sleep(10); 1; }’):
```

which implies that the internal state of the `future` seems corrupted somehow.