Assume columns can have missing values by default

[cynddl]

The current behavior of CSV.jl is to scan columns until row_for_type_detect iterations have passed and infer their type.

However, for large files with few null cells, this force to set row_for_type_detect to the maximum number of rows, as the heuristic cannot detect if the last row of the file might contain null cells for instance.

---

Solution

One workaround would be to assume that rows can contain null cells by default, except if the user disable it. This was the behavior in DataFrames.readtable for instance.

Another would be to have a manual flag in CSV.Options to assume that all rows can contain null values.

[nalimilan]

I think nullable should accept three values, e.g. :all, :auto, :none. In the long term it would be nice to widen the column type as soon as a missing value is detected (due to the way arrays with missing values are stored, it should be possible to do that without copying already parsed values).

[nalimilan]

After discussing this with @quinnj, we agreed that assuming columns can contain missing values was the right default. I've prepared #173 to change this.