Make nodata casting consistent based on new array type

[rhugonnet]

This PR fixes the fact that set_nodata was circumvented during from_array() creation, making the behaviour fully consistent for checking nodata input and setting.
Additionally, for practicality, it adds a cast_nodata boolean argument to from_array() defaulting to True which re-casts the input nodata to the default of the array type, if the nodata value was incompatible with the input array type.

The impact of this last change was actually much less of a bigger than I thought, because all integer nodata values are compatible with floating array types and the other direction rarely happens. There is almost no NumPy operations that convert from float to int (even operations like floordiv keep the output a float even if the output of the operation could technically be integers without loss of information, for instance).

So this mostly addresses the cases where the user is copying the data with a new array that has different data type (fairly common), or provided a non-compatible nodata value directly to from_array (less common).

Resolves #517
Resolves #516

TO-DO:

• Use set_nodata() in from_array section of __init__(),
• Add optional casting of nodata for new array type in Raster.copy(),
• Add automatic casting of nodata for NumPy operations.
• Add tests of new behaviour.

[adehecq]

All good ! There's just a block of code that was deleted but I don't understand why.