feat(llamacpp): allow verified estimator overruns

[sjug]

Summary

• Allow a custom llama.cpp relaunch policy to attempt its exact configuration when a valid estimator result predicts less free VRAM than the requested target.
• Persist the permission separately from whether the bypass was actually used, preserving auditability across retries and transactional rollback.
• Add the custom-only runtime-editor checkbox, expose verified estimator status, and align the KV cache selector with the neighboring unit-backed controls.

Safety

• Automatic sizing remains estimator-strict, and malformed output or estimator execution failures cannot be bypassed.
• Static context limits, full GPU offload, and measured post-load free VRAM remain mandatory.
• Failed launch or verification attempts use the existing transactional rollback to restore the previous persisted policy.

Testing

• Python 3.12: 1,495 passed
• Python 3.13: 1,495 passed
• Coverage: 92.59%
• mypy: 143 files, no errors
• ruff check, ruff format, shellcheck, and git diff --check: clean