Multithreading is not used for IDCT if the rayon feature is enabled

[Shnatsel]

hyperfine shows the image decoding being significantly faster with the rayon feature disabled:

Benchmark 1: jpeg-bench/target/release/jpeg-bench "some-large-jpeg.jpg"
  Time (mean ± σ):     705.5 ms ±  22.7 ms    [User: 830.4 ms, System: 151.0 ms]
  Range (min … max):   680.5 ms … 742.2 ms    10 runs
 
Benchmark 2: jpeg-bench-parallel/target/release/jpeg-bench-parallel "some-large-jpeg.jpg"
  Time (mean ± σ):     867.2 ms ±  11.1 ms    [User: 906.7 ms, System: 120.5 ms]
  Range (min … max):   854.6 ms … 888.7 ms    10 runs
 
Summary
  'jpeg-bench/target/release/jpeg-bench "some-large-jpeg.jpg"' ran
    1.23 ± 0.04 times faster than 'jpeg-bench-parallel/target/release/jpeg-bench-parallel "some-large-jpeg.jpg"'

Looking at the profile, it seems that the additional threads for IDCT are spawned only when the rayon feature is disabled (profile). When it is enabled, the IDCT is happening on the main thread (profile).