Add self-contained AudioLDM2 Large inference

[ml-maple-monk]

Summary

• add an Understand Anything architecture and inference-path cache
• vendor the AudioLDM2 source byte-for-byte at commit b5786c5
• add pinned artifact resolution for audioldm2-full-large-1150k and its tokenizer/config assets
• expose one public self-contained/run_inference.py command with strict checkpoint loading
• document the 64-bin log-mel KL VAE and 16 kHz HiFi-GAN codec path
• add fast contracts and an opt-in real checkpoint-to-WAV acceptance test

Verification

• default suite: 17 passed, 1 opt-in E2E skipped
• real E2E: 1 passed in 774.37 seconds on CUDA
• checkpoint: 11,470,645,003 bytes; SHA-256 6b483a47480d15d90c2eaaa17415484538bff7c6d9d2282f7838ad004c637c09
• output: mono 16 kHz, 163,872 frames, finite and non-silent
• vendored source and license parity tests pass against Git object b5786c5

The generated checkpoint cache, isolated environment, and proof WAV remain ignored.