train_imdb: coredump

Hi dynet-users,

I'm trying train the imdb example with the following input data (small entry):

```text
5 ||| a traveller experiences strange paranormal activity in a room at an inn where he stops to stay the night . | clothes fly around , a chair vanishes and his boots walk off ; all to the consternation of the spooked out man . | this little comic short from georges méliès is a very early showcase for his cinematic trickery and visual invention . | just as significantly it illustrates his sense of humour . | méliès made many funny films which incredibly are still amusing even now , over a century down the line . | the effects remain impressive even though we all know how they are done . | there is always a loving care in méliès work . | this one is no exception and shows why he is the first cinematic genius .
```

I'm not quite 100% sure, if this is the correct input format, but I'm planning to write a small documentation for the `imdb` example.

Here's the output:

```bash
[stefan@arch-mse examples]$ ./train_imdb ~/imdb_dataset/imdb-train.txt ~/imdb_dataset/imdb-dev.txt
[dynet] random seed: 763958800
[dynet] allocating memory: 512MB
[dynet] memory allocation done.
Reading training data from /home/stefan/imdb_dataset/imdb-train.txt...
280593 lines, 3929753 sentences, 87596427 tokens, 115353 types
Labels: 10
Reading dev data from /home/stefan/imdb_dataset/imdb-dev.txt...
33793 lines, 475459 sentences, 10599895 tokens
Parameters will be written to: imdb_100_200_400_200-pid3185.params
**SHUFFLE
[lr=0.001 clips=20 updates=100] E = 2.35599 ppl=10.5486 (acc=0.13) NORM: 350.165
[completed in 10731.7 ms]
[lr=0.001 clips=32 updates=100] E = 2.42035 ppl=11.2498 (acc=0.15) NORM: 350.542
[completed in 11682.4 ms]
*** Error in `./train_imdb': corrupted size vs. prev_size: 0x0000558d99610720 ***
```

I'm using 6e1e9f496973c7f77d85aea44481e9a50c654249 of `dynet`, `eigen` 3.3.4, and GCC 7.2 on Arch Linux.