hifi-decode: Performance enhancements, Fix head switching interpolation

[eshaz]

• Fixed peak detection issue with head switching interpolation.
  • Hifi-decode processes audio in blocks, and each of these blocks have an overlap to remove an amplitude spike that occurs at the edges of each block. Head switching noise reduction wasn't filtering all of these spikes and this was interfering with the peak detection.
• Fixed possible divide by zero error in auto fine tune mode
  • Limited the amount of carrier adjustment, preventing it from going to zero.
• Logging features
  • Logging updates more frequently on each block read and audio output.
  • Logging shows audio latency which is helpful for preview mode / real time rf previewing
• Use shared memory to transfer data throughout the decode pipeline
  • Significantly speeds up soundfile decoding
  • Avoid unnecessary memory copying where possible
  • When Saving resampled data to out pointer instead of returning a copy tuxu/python-samplerate#26 is addressed, additional memory copies can be avoided
• More usage of Numba optimizations
  • Optimized FM demodulation code
  • Installed rocket-fft to allow Numba fft optimizations.
  • Installed icc_rt for Intel vector intrinsic support for Numba
• Better handling of the last block so all the output data is flushed at the end.
• Simplify decoder pipeline. Each process is connected directly together with queues.
• Removed single threaded decode to simplify the code
• Better documentation in the code