Fix Higgs Audio codec decode seams

[0xShug0]

Summary

Fixes #429.

Higgs Audio TTS used a small left-only context when chunking codec decode for generated code sequences longer than the full-decode limit. The codec decoder has a wider convolutional receptive field, so hard-concatenating chunks with only 8 frames of left context could leave audible clicks at chunk boundaries.

This changes the chunked decode path to borrow 32 frames of context on both sides of each emitted window, then trim back to the center emitted frames.

Validation

Built audiocpp_cli in debug mode with CUDA and ran the two issue repro cases using the same seed/reference/text pattern from the report.

Case	Before seam jump/rms	After seam jump/rms	Result
English, 10.24s boundary	1.808	0.039	fixed
English, 20.48s boundary	1.014	0.089	fixed
Chinese, 10.24s boundary	0.555	0.029	fixed
Chinese, 20.48s boundary	1.471	0.213	improved

Runtime cost from the larger context window:

Case	Before RTF	After RTF	Change
English	0.1156	0.1161	+0.4%
Chinese	0.1201	0.1218	+1.4%

The cost only applies to the chunked codec decode path for longer generated outputs; short outputs that fit the full-decode path are unchanged.