refactor: centralize runner weight staging and cleanup

[leejet]

Summary

• Add a shared runtime weight manager for lazy parameter loading, staging to compute backends, LoRA application, and release.
• Wire runners, conditioners, extensions, VAE/audio VAE/control paths, and graph-cut execution through the new manager and runner_done() lifecycle.
• Split compute cleanup controls into auto_free, free_compute_buffer, and free_compute_params.
• Update graph-cut max-vram merging/residency handling and switch model-loading progress fill to #.

Related Issue / Discussion

N/A

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.