In-place guarantees for scan algorithms

[gevtushenko]

This PR addresses multiple issues of in-place execution for scan algorithms:

Scan:

• LOAD_LDG is removed from tuning policies to guarantee in-place execution. Affected architectures are Kepler and Maxwell. Benchmarks showed no difference in performance on those platforms.
• In-place execution guarantees are highlighted in the documentation of the existing API
• In-place API is provided
• In-place tests are fixed. The code used to compare decltype(*d_in) with OutputT which always led to comparing T& with T. Therefore, we've never tested in-place execution.

Scan by key:

• LOAD_LDG is removed from tuning policies to guarantee in-place execution.
• To preserve thrust documentation guarantees, implementation is changed to allow aliasing of keys/results. This change affects performance on small to moderate problem sizes (up to 6% slowdown).
• In-place execution guarantees are highlighted in the documentation of the existing API
• Tests of in/out values aliasing are fixed. The code used to compare decltype(*d_values_in) with OutputT which always led to comparing T& with T. Therefore, we've never tested in-place execution.
• Tests of keys/values aliasing are added.

[alliepiper]

Some minor issues in doc strings, but otherwise LGTM.