Assemble detector modules without copying them into a 'stack' array

[takluyver]

From discussion relating to KaraboFAI: if you want to load data from AGIPD/LPD/DSSC and assemble it, it currently gets copied twice in memory:

individual module arrays -> stacked array -> assembled image

The first of these copies should be unnecessary. There are two basic schemes for how we could avoid it:

• Make a 'virtual' stack, e.g. a list of arrays, to represent the data without copying. We don't need to do any maths between the modules, so they don't need to be in a single array.
• Read the data directly into slices of the stacked array, so we don't allocate individual arrays for each module.

[zhujun98]

I would go for the first method and make a StackedData class which has several methods to provide the necessary np.array interface, e.g. shape property:

1. The stacked data provides the similar interface as the data from the calibration pipeline, although one needs to move the axis twice for LPD and DSSC. Therefore, it is good to 'keep' it;
2. Depending on the level of backward compatibility you want to have, there is a few options to make the new stack_detector_data function. No matter what, you do not need to touch the method position_modules_fast which is used for assembling.

[takluyver]

PR #196 is an initial attempt at implementing option 1.

[takluyver]

Done in #196.