fix scalefactor delta overflows to comply with AAC spec (ISO 14496-3)

[nschimme]

This commit addresses two bugs in BlocQuant()'s scale factor range pass that allowed scalefactor differences to exceed the strict ±60 limit required by the AAC specification:

1. Fix PNS delta predictor: PNS scale factors share the HCB_PNS codebook and require a separate delta predictor. Without it, the first PNS band's delta was computed relative to the regular global_gain, producing out-of-bounds deltas. A dedicated lastpns predictor is now initialized 90 steps below global_gain so the first PNS entry fits comfortably within the ±60 constraint.
2. Enforce limits on all active bands: The quantizer previously only clamped deltas for HCB_ESC bands. The condition is now updated to (book != HCB_ZERO) && (book != HCB_NONE) to ensure the ±60 limit is enforced for every active scalefactor band, regardless of the Huffman codebook.

Additionally, this refactors the clamping logic into a centralized clamp_sf_diff() inline function and replaces hardcoded scalefactor magic numbers with named constants in huff2.h.

This causes a slight regression to speech where they accidentally did better, but overall all other cases we see a slight MOS improvement at the cost of 1% CPU throughput that we need to pay anyway for correctness: https://github.com/nschimme/faac/actions/runs/24264165540

[nschimme]

Context on why SF_PNS_OFFSET was 90:

The AAC Scalefactor Delta Bug

The Bug: "The Broken Bridge"

AAC uses Huffman Book 12 to encode the difference (delta) between band volumes.

• The Spec: Limits this delta to exactly ±60.
• The Error: The original code didn't enforce this for all bands and used one "memory" (lastsf) for both music and noise (PNS).

The Scenario

1. Music Band: Volume = 140.
2. Noise Band (PNS): Volume = 50.
3. Calculation: $50 - 140 = -90$.
4. Result: $-90$ is outside the allowed ±60. The encoder writes an "out-of-bounds" index, making the file unplayable or "glitchy" on standard players.

The Fix: "Two Tracks & Guardrails"

• The Guardrail: clamp_sf_diff ensures no delta ever exceeds ±60, keeping the bitstream legal.
• The Second Track: By adding lastpns, noise bands now have their own "memory" separate from music.
• The "90" Offset: By starting the noise memory 90 steps below the music, you ensure that quiet noise levels can be reached in a single legal step.

Why Speech MOS Dropped

In speech, "noise" (consonants like S or F) is often louder than background hiss.

• If the consonant is only 30 steps below the music, but your code starts looking 90 steps below, it has to "climb up" to reach the right volume.
• If it can't climb fast enough due to the ±60 limit, the consonant loses energy, sounding "muffled" or "lisp-like."

[nschimme]

Okay, I discovered that this is partially responsible for #40. I'm working on the remaining fix.

[nschimme]

Okay, this fixes #40 and possibly other cases too. No performance drop: https://github.com/nschimme/faac/actions/runs/24285523096

[nschimme]

Hopefully the above two fixes also let me finally got Pseudo SBR to work. I kept getting blocked 😅