learner$train() informative error for incorrect sampler

[tdhock]

Closes #420

This PR affects this code example #420 (comment)

Before this PR we get

> inst_learner$train(sonar_task)
Error in ctx$batch$y$to(device = ctx$device) : 
  attempt to apply non-function

After this PR we get a more informative error which should be able to help the user do a fix:

Error in train_loop(ctx, callbacks) (from mlr3torch-bug.R#55) : 
  dataloader_next() returned NULL, which means there are no more samples/batches. Typically this occurs when length of sampler/batch_sampler is greater than the number of samples/batches. Please modify .length() method to return the correct number (samples for sampler, batches for batch_sampler), which should be equal to the number of times that .iter() can be called before returning coro::exhausted()