Is there any way to do discrete masking?

I would like to implement a discrete mask like the following one in numpy.
```python
import numpy
a = numpy.array([0., 2.])
a = 1/a
a # array([inf, 0.5])
b = numpy.array([True, False])
a[b] = 0
a # array([0. , 0.5])
```

If we use an ordinary mask implementation, for example, setting `b = input(cg, {2}, {0., 1.});` and `a = cmult(a, b); // a is an Expression just like the one in 4th line of the code block above`, then `nan` values would appear. The reason is that 0 * inf is always indetermined.

Could we implement such discrete masks on top of current dynet or it is a missing feature?

Start with the current DyNet paths for input and cmult mentioned in the issue, and compare their behavior with the NumPy example. Determine how a discrete mask could avoid evaluating the masked 0 * inf operation; done means the mask produces the intended values without nan in this case.

Written by the indexing model from the issue text.