Automate manual audio regression comparisons

[derselbst]

The manual rendering tests previously required human inspection and could not be executed in CI. Inspired by the discussion in #1747, I fired up copilot to introduce an automated workflow that compiles a reference FluidSynth revision, renders the audio with the reference and current version of FluidSynth, and compares SNR/RMS/ABS metrics, without requiring versioned "golden reference" files.

• Allow check_rendering to run against a caller-specified fluidsynth binary via MANUAL_TEST_FLUIDSYNTH.
• Add test/run-manual-regression.sh to build current + reference revisions, render both sets, and compare with SoX metrics and thresholds.
• New Linux workflow rendering-regression.yml to run the regression comparison with configurable reference ref and thresholds.

Example usage:

REFERENCE_REF=HEAD~1 SNR_MIN=60 RMS_MAX=0.0001 ABS_MAX=0.01 test/run-manual-regression.sh

TODO:

• test suite currently fails when and old version of fluidsynth is being passed new / unknown commandline options, like -o synth.limiter.active=1
• SNR, RMS and abs diff need to be evaluated to get meaningful values

[derselbst]

@spessasus FYI

[derselbst]

Why did I close this?

[spessasus]

I have no idea! The PR looks good so far.