Phosphor performance tuning

For performance improvements that have been suggested or considered but are
currently premature:

• Check enabled flag with relaxed atomic
• Try avoiding the conditional branch caused by the ternary expression in
  TraceLog::logEvent
• Use __builtin_expect where useful
• Sentinel benchmarking suggests that double the number of logical cores
  is the sweet-spot for the size of a shared array of ChunkTenants.
• Reduce size of trace event by reducing to 24 bits of microseconds (~16 seconds
  from 2^64 nanoseconds and storing a further 16 bits of resolution in the chunk
  header. (Saves ~4 bytes since type can be combined into timestamp)
• Reduce size of trace event by combining trace category and trace name pointers
  (Saves 8 bytes as its one less pointer)
• Assume one less argument on Async events and use id as one of the arguments
  (Saves 8 bytes as there's no need for separate id property)
• Add a duration property (Costs 24 bits)