DelayedConvert feature - #1797

This feature solves a very specific problem: Sometimes you want to map over a Tile, or combine a MulitbandTile's bands, or the like, and have the resulting Tile be of a different CellType then the original tile. For instance, when doing NDVI, if you do this on a IntConstantNoDataCellType multiband raster:

multibandTile.combineDouble(3, 4) { (r, ir) =>
  (ir - r) / (ir + r)
}

You'll get garbage values, since the ratio produces results from [-1, 1], but our result tile will be IntConstantNoDataCelltype, which will not hold the floating point values. What you have to do now is convert it to a floating point valued raster first:

multibandTile.convert(DoubleConstantNoDataCellType).combineDouble(3, 4) { (r, ir) =>
  (ir - r) / (ir + r)
}

This is awful for performance, because not only do you have to iterate over band 3 and 4 to set each value to a double value before you do the combine, but this will actually iterate over bands you don't even care about. What you can do to get around the latter issue is this:

multibandTile.subsetBands(3, 4).convert(DoubleConstantNoDataCellType).combineDouble(0, 1) { (r, ir) =>
  (ir - r) / (ir + r)
}

but that doesn't not solve the latter problem.

This solves this particular brand of problem by creating a delayedConversion method, which returns a Tile or MultibandTile that is exactly like the parent tile, except that for any method that produces a Tile or MultibandTile, that resulting tile will be an ArrayTile or ArrayMultibandTile with the target cell type. Our NDVI calculation now becomes:

multibandTile.delayedConversion(DoubleConstantNoDataCellType).combineDouble(3, 4) { (r, ir) =>
  (ir - r) / (ir + r)
}

And now we do the thing we want: iterate over band 3 and 4 only, and set the double NDVI value into a DoubleConstantNoDataCellType.

NOTE

Also contained in this PR is some ArrayMultibandTile optimizations.

Connects #1410