Big refactoring to the tuner link

• Now user friendly
• All dependencies installed via pip (except from profilers)
• Automatic GPU vendor detection
• Possibility to add a time budget to indicate a desired duration of the tuning
• Biggest change: now independent steps (ensemble of kernels) are tuned togheter using multiple optuna studies at the same time
  • Single run of the standalone benchmark
  • Profiling of multiple step with the same run
  • One optuna study per step, which will suggest a new configuration
• This heavily reduces the time needed for tuning