Minimize thrust dependency

[stotko]

In order to implement several functions in a portable manner across the backends, thrust is used as the abstraction of choice. This may, however, limit extending support to other possible libraries making it desirable to reduce the coupling to thrust. Throughout the whole code base, the following classes and functions are used:

Header algorithm:

• all_of
• copy
• copy_if
• count_if
• equal
• fill
• for_each: Only used in stdgpu::for_each_index implementation.
• generate
• reduce
• sequence
• sort
• tabulate
• transform
• transform_reduce: Only used in stdgpu::transform_reduce_index implementation.

Header execution:

• device: Only used as object alias for stdgpu::execution::device.
• host: Only used as object alias for stdgpu::execution::host.

Header functional:

• equal_to<>
• identity<>
• plus<>

Header iterator:

• counting_iterator: Only used in stdgpu::for_each_index and stdgpu::transform_reduce_index implementation.
• discard_iterator: Implementation detail used for iterator_adaptor.
• distance
• iterator_adaptor: Only used in implementation of thrust-compatible container iterators.
• transform_iterator: Only used in stdgpu::transform_range implementation.
• zip_iterator

Header utility:

• pair: Only used as typedef for stdgpu::pair.
• tuple

Other Headers:

• random

Note: The examples may still make use of thrust to demonstrate interoperability.

[stotko]

The remaining places, where thrust is used, are listed behind the respective elements. Any further replacements require significantly more effort than the so-far completed tasks. Therefore, the current level of coupling is sufficiently low for now and could be revisited again in the future if required.