Fix post-echo artifacts by pre-clamping scalefactor deltas (Fixes #40)

[nschimme]

This PR eliminates the post-echo artifacts that occur in quiet bands immediately following loud bands. It resolves an architectural bug where the encoder and decoder were applying different gain values due to delayed scalefactor delta clamping.

Fixes #40

Architectural Changes

• Pre-Quantization Clamping: The post-loop delta clamping in BlocQuant() modified sf[] after qlevel() had already quantized the spectral data. When a negative delta exceeded -SF_DELTA, the decoder reconstructed using the clamped (larger) value while the encoder used the smaller (natural) gain. Moving the delta clamp inside qlevel()—prior to quantization—ensures strict mathematical synchronization between the bitstream and physical acoustics.
• Refactored Gain Extraction: The mathematical clamping and Huffman overflow limits have been extracted into a single gain_with_overflow_clamp() helper to prevent peak clipping when the gain is artificially boosted by the clamp.
• Preserved IS Stereo Biases: Variable names in qlevel() have been updated (sf_rel, sf_abs) to explicitly distinguish between relative bitstream values and absolute energy levels. This ensures the pre-loaded IS stereo biases from AACstereo() are correctly tracked without corrupting the delta chain.
• global_gain Simplification: Migrating the delta chain into qlevel() allowed for the cleanup of the bottom half of BlocQuant(). The global_gain is now cleanly derived from the first active band sequentially.

Verification

• 128 kbps: Frames 17-20 amplitude ratios are 1.028/0.984/1.000/1.075 (within normal codec loss; post-echo eliminated).
• q=5000 (Perfect Reconstruction): All frames remain within 0.995-1.005 of the original.
• Benchmark passes at 99.8% consistency and no performance regressions.

decoded-preclamp.wav.zip

[fabiangreffrath]

I have to admit that I don't understand most of the changes introduced here, but I trust you (and Claude 😉) and the results enough to just wave this through.

[nschimme]

We have a rather epic 100 count PR for this too