pns: fix decoding window offset clamping for short sequences

[nschimme]

Fixes a bug in pns_decode() where Perceptual Noise Substitution (PNS) band bounds (begin/end) were incorrectly clamped against ics->swb_offset_max instead of the full buffer size (frame_len).

For EIGHT_SHORT_SEQUENCE frames, swb_offset_max spans only a single short window (frame_len / 8). As a result, base = group * nshort evaluated to >= swb_offset_max for every window group after the first. Clamping against swb_offset_max caused min(base + swb_offset[sfb], swb_offset_max) to collapse to swb_offset_max, resulting in begin == end and silently skipping PNS noise generation across 7 of the 8 window groups in short frames.

Fix: Clamp band offsets against frame_len (already passed into pns_decode()) to reflect the actual bounds of the spectral array.

Fixes #191

---

Verification & Benchmark Results

Tested on a low-bitrate FAAC-encoded AAC-LC stream (32kbps, PNS enabled, using short-window block switching):

Metric	faad2 (before)	faad2 (after)	ffmpeg (ref)	Apple afconvert (ref)
>8kHz Energy Ratio vs. Source	1.52x	3.57x	3.57x	3.48x
Zimtohrli MOS	4.0887	4.5640	4.5519	4.5644

• Decoder parity: After the fix, high-frequency energy alignment and perceptual quality (MOS) match reference decoders (ffmpeg and Apple afconvert).
• Control check: Disabling PNS (--pns 0) yields identical PCM output between faad2 and ffmpeg.
• Regressions: Gapless sample counting across the regression test suite remains unaffected.

---

Test Plan

• Build succeeds clean with CMake/Ninja.
• Verified high-frequency noise spectral energy matches reference implementations.
• Verified Zimtohrli MOS improvement on PNS-heavy short-window test streams.
• Verified existing regression suite passes without sample count drift.