Summary

• Adds targeted single-device refresh methods on AllSmi: get_gpu_by_uuid / refresh_gpu, get_cpu_by_index / refresh_cpu, get_memory_by_index / refresh_memory. Each refresh_* returns bool and leaves the original struct untouched on a miss.
• Adds a 0-based stable index: u32 field on CpuInfo and MemoryInfo, populated by AllSmi while flattening readers, marked #[serde(default)] for snapshot and wire backward compatibility. CPU topology is static so the index is a stable correlation key across refreshes; MemoryInfo.index exists for API symmetry.
• Adds a default trait method GpuReader::get_gpu_info_by_uuid (filters the full enumeration). The NVIDIA reader keeps the default for this PR per the issue's staged plan; an NVML device_by_uuid fast path is the logical follow-up.
• Extends examples/library_usage.rs with a correlate-and-refresh loop covering GPU, CPU, and memory.

Why this approach

• New methods live on AllSmi, not on the info structs. GpuInfo / CpuInfo / MemoryInfo are plain Serialize + Deserialize + Clone DTOs with no hardware handle; the readers already live inside AllSmi, so that is the correct owner of refresh logic. A self-updating update(&mut self) on the structs would have to embed a #[serde(skip)] reader handle that is None for any deserialized value.
• GPUs key on uuid (already unique per device, hot-plug / MIG reconfiguration safe). CPUs and memory key on index because their topology is static and there was previously no per-entry identifier.
• All new methods take &self; AllSmi: Send + Sync is preserved.
• Additive surface — only consumers that build CpuInfo / MemoryInfo via exhaustive struct literals see a compile break. Internal call sites are updated; downstream consumers normally receive these structs from the library rather than building them.

Test plan

• cargo fmt --check
• cargo clippy --lib --tests -- -D warnings
• cargo check --lib --tests
• cargo test --doc client (14 doctests, including the new get_gpu_by_uuid and refresh_gpu examples)
• cargo test --test library_api_test — 25 tests pass, including the 9 new tests for by-uuid hit/miss, in-place refresh idempotency, index stability across consecutive calls, and out-of-range index handling
• cargo test --test snapshot_test — confirms index round-trips through JSON
• cargo run --example library_usage — prints fresh utilization for one GPU by UUID, one CPU by index, and memory by index on a system with one NVIDIA GB10 GPU
• View-cache and renderer unit tests still pass after the struct-literal updates

Follow-up

• NVIDIA device_by_uuid fast-path override on NvidiaGpuReader::get_gpu_info_by_uuid (the static cache is currently keyed by index; a uuid resolution layer or a uuid-keyed static cache will make this a single-NVML-handle path instead of filtering the full enumeration).

Closes #211