Added 3 different modes for direct lake models so that we do not always force all columns to be transcoded into memory.

• ResidentOnly - only run the distinctcount queries on columns that are current resident in memory
• Referenced - run distinctcount queries on columns that are referenced by measures or relationships. This should provide the minimum information required for performance analysis of measures and for tools like DAX Optimizer.
• Full - run distinctcount queries on all columns

I've added this option since the default behaviour at the moment is the same as Full and this can be detrimental to large models. It can take longer than necessary and potentially may still not produce the desired result since as we cycle through all the columns it could force some of the earlier columns to be paged out of memory.

I've set ResidentOnly as the default as it's the cheapest/fastest option.

Note that if an option other than ResidentOnly is chosen I am re-running the DMV collection after the stats collection has been run since the memory usage is usually changed as a result of running the distinctcount queries.

I did consider whether to add the selected Direct Lake analysis mode as a property to the vpax file, but I decided against his for the time being since it can be implied by checking the IsResident property on a column.