fix: vGPU borrow-ergonomics follow-ups from #130

[brayniac]

Follow-ups to the review discussion on #130 — no behavior changes, all API-loosening:

• VgpuInstance::instance_type() now takes &self instead of &'dev self. The old receiver borrowed the instance for 'dev, so keeping the returned VgpuType beyond the instance's scope (e.g. querying the type, dropping the instance, then using the type) failed with E0597: does not live long enough. (Patterns where the VgpuType stayed inside the instance's scope already compiled, since covariance let 'dev shrink at the call site.)
• Device::active_vgpus() drops the unnecessary <'a> ... where 'a: 'nvml clause and returns VgpuInstance<'_> tied to the device borrow. The old clause forced 'a == 'nvml; the new signature accepts a strict superset of caller patterns (Device is covariant in 'nvml).
• VgpuInstance::new is now public with docs, mirroring VgpuType::new. This gives non-Linux builds a constructor story and removes the dead_code warning there, where the Linux-gated active_vgpus() was its only caller. A garbage instance ID is safe: every FFI path checks the NVML return status, so it yields Err(NotFound)/Err(InvalidArg), never UB.

Verified: compile probes for the fixed patterns pass on Linux and macOS targets (and reproduce E0597 against main as a negative control); cargo check --all-features (lib + tests) is warning-free on both; cargo fmt --check passes; clippy and cargo doc show only pre-existing warnings. No post-1.60 constructs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FVPWxFN8rioYb8JhreYQmJ