cudf::column redesign

[jrhemstad]

Closes: #1443

Depends on rapidsai/rmm#99

In effort to provide a replacement for gdf_column and related infrastructure, this PR introduces a cudf::column and associated classes.

Goals:

• Memory allocation is handled via constructors
• Memory ownership is well-defined
• Control shallow vs. deep copy w/ C++ move semantics
• Access to data and bitmask is controlled via member functions
• All memory allocated by libcudf is owned by a C++ object

Non Goals:

• Implement types other than primitive, fixed width types (e.g., strings, categories, etc.)
  • But we will provide a path forward for these types

What's new?

• cudf::column
  • Memory owning class
  • Allocates memory on construction, deallocates on destruction
  • Enables compound column types via a vector of "child" columns
• cudf::column_view
  • Non-owning, immutable class allows viewing data as a column
  • Similar to a llvm::ArrayRef can be constructed from a cudf::column or a void* to device memory
  • All column inputs to libcudf will be as column_views
• cudf::column_view
  • Non-owning, mutable class allows viewing data as a column
  • Similar to a llvm::MutableArrayRef, can be constructed from a cudf::column or a void* to device memory
  • All column in/out arguments to libcudf will be as mutable_column_views
• cudf::column_device_view
  • Same as above, but with __device__ accessors
• cudf::mutable_column_view
  • A mutable column_view, for in/out parameters to libcudf functions
• cudf::exp::table
  • Owning set of cudf::columns of the same size
  • exp namespace to disambiguate from existing cudf::table
• cudf::table_view
  • Set of column_views of the same size
• cudf::mutable_table_view
  • Set of mutable_column_views of the same size

[raydouglass]

rerun tests

[kkraus14]

Test failure is innocuous, we have a 10 decimal point equality check for a rounding function in number and it was equal to 9 points, first time seeing that failure. Will admin merge this once the last gpu tester comes back and follow up on that test in an issue / PR to reduce the precision.