Update README and CHANGELOG for 1.15.0-rc0.

Fix CachingDeviceAllocator with debug=true

Test that <algorithm> and <memory> aren't included directly.

New segmented sort algorithm

These headers cause circular dependencies when Thrust is used as
a backend to nvc++'s stdpar. They've been used accidentally several
times, so it's probably worth automating a check for them.

Expanded and renamed the namespace macro check to test for these headers.

Avoid include cycles with nvc++ -stdpar

An `#include <algorithm>` in cub/device/dispatch/dispatch_scan_by_key.cuh
was causing compilation failures due to an include cycle when using
`nvc++ -stdpar`.  It doesn't look like `<algorithm>` is needed at all in
this file, so I am simply removing it.

Add scan by key algorithms to cub

Adds BlockRunLengthDecode algorithm and tests