Multi-benchmark tuning

[gevtushenko]

Description

closes #151

This PR provides several CUB tuning infrastructure improvements. It allows multiple benchmarks in a single tuning, provides a new workload to batched memcpy benchmark, and combines all segmented sort benchmarks in a single tuning. Besides that, brute force seeker now randomizes variant space to improve coverage. Benchmark is now killed when the search is terminated. All runtime workloads are now covered in a single benchmark launch, reducing the overhead of context initialization. This optimization provides about 20% speedup of one generation with five workloads per benchmark. This number should be more significant for benchmarks containing many workloads (segmented algorithms).

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[elstehle]

Approving for benchmark changes 👍