Improve frame_sync_impl::get_symbol_val() performance

[mskvortsov]

On Apple M3 hyperfine shows 22% of improvement:

Benchmark 1: python3 gr_meshtastic.py
  Time (mean ± σ):     428.8 ms ±   6.7 ms    [User: 420.3 ms, System: 81.5 ms]
  Range (min … max):   420.8 ms … 439.1 ms    10 runs

Benchmark 1: python3 gr_meshtastic.py
  Time (mean ± σ):     335.1 ms ±   1.6 ms    [User: 326.2 ms, System: 79.8 ms]
  Range (min … max):   332.7 ms … 337.0 ms    10 runs

On RPi4 the improvement is about 13% (8.12s to 7.09s). Some details on the benchmarking setup are in #92.

There are other places in the code where a state returned by kiss_fft_alloc() is not reused, but it seems that they are not on the hot path.

[mskvortsov]

Also fixed minor imperfections found by valgrind.

[tapparelj]

Hello @mskvortsov ,
thanks for the fix, all looks fine to me!
There is an issue with the build of the conda package so I'll fix it and merge it when I'm done.