Remove runtime aerialvision overhead w/ preprocessor

[CoffeeBeforeArch]

Despite the AerialVision logging not being enabled in a GPGPU-Sim config, you still pay the price of runtime overhead for collecting memory system stats on each window. This commit seeks to remove most of this overhead by adding GPGPU-Sim build mode (av_enabled). When av_enabled is not enabled, the preprocessor removes much of the code related to AerialVision in the caches.

Below are examples of where the overhead of AerialVision is paid even when it is not enabled:

• The cache_stats constructor in gpu-cache.cc
  • A vector of vectors is created for per window stats
• the cache_stats operator+=
  • cache_stats_pw is updated

This lead to an observed 3-5% increase in observed performance.