fix(fish-speech): support CUDA 13 on arm64

[localai-org-maint-bot]

Description

This PR fixes #11344.

• expose the cloned fish-speech source through PYTHONPATH after backend relocation
• pin CUDA 13 builds to PyTorch 2.9.1 CUDA wheels and align fish-speech upstream metadata before editable installation
• decode reference audio with soundfile, avoiding torchcodec, which has no Linux aarch64 wheels
• keep upstream preparation idempotent and fail the image build if the expected upstream source shape drifts

The partial-cuDNN failure reported in the same issue is already guarded on master by #10946; its existing regression suite is included in verification rather than duplicating that logic here.

Notes for Reviewers

Verified with:

• make -C backend/python/fish-speech test-unit
• bash scripts/build/package-gpu-libs-cudnn_test.sh (14 cases)
• Bash syntax checks and Python byte-compilation
• two consecutive preparation passes against a fresh clone of fishaudio/fish-speech main
• git diff --check for both LocalAI and the transformed upstream tree

A full CUDA 13/aarch64 backend image build was not run on this host.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable