Reduce Vevo2 FM graph peak memory

[0xShug0]

Fixes #437

Summary

• split Vevo2 FM step mutable IO tensors from compute graph temporaries
• allocate the FM step compute graph with ggml gallocr/liveness allocation instead of allocating every tensor in the graph context
• release transient FM graph resources after conditioning and final mel read

Validation

Case	Result	Peak VRAM
pre-fix C++ VC 60s + b.wav	failed: requested 82460.49 MiB CUDA allocation	n/a
pre-fix C++ SVC 60s + b.wav	failed: requested 82460.49 MiB CUDA allocation	n/a
after-fix C++ VC 60s + b.wav	completed	5289 MiB
after-fix C++ SVC 60s + b.wav	completed	5295 MiB
Vevo2 CUDA path tests	4/4 passed, 0 artifact differences vs baseline	SVC path 15418 -> 6346 MiB

The 60s validation used a 60-second concat of bundled Vevo2 gallery audio as source and the repo b.wav asset as the reference voice.