JSON datashape discovery doesn't recognize NaN as Option type

With an example CSV file with missing fields like:

```
textcol,numcol
abc,123
,456
def,
```

odo.discover correctly identifies that the columns should be Option dtypes (i.e. the columns can contain missing data):

```
>>> import odo
>>> csvfile = odo.CSV('odo_example.csv', has_header=True)
>>> odo.discover(csvfile)
dshape("var * {textcol: ?string, numcol: ?float64}")
```

When we use odo to convert this CSV to a JSON file, the JSON missing values are represented by NaN, which I don't think is entirely within the JSON spec, but the Python json package serializes it that way and I think it's reasonable. So running this code...

```
>>> jsonfile = odo.JSON('odo_example.json')
>>> odo.odo(csvfile, jsonfile)

```

...produces the following JSON data:

```
[{"numcol": 123.0, "textcol": "abc"}, {"numcol": 456.0, "textcol": NaN}, {"numcol": NaN, "textcol": "def"}]
```

However, running odo.discover on the resulting JSON file with NaN representations of missing data doesn't result in Option dtypes.

```
>>> odo.discover(jsonfile)
dshape("3 * {numcol: float64, textcol: string}")
```

I think JSON discovery should recognize NaN as missing values and set columns with NaN data as Option dtypes.