Summary

Completes the label standardization started in PR #176 for NVIDIA GPUs. All non-NVIDIA NPU exporters (Tenstorrent, Rebellions, Furiosa, Intel Gaudi, Google TPU) now emit npu_index/npu_uuid in place of the generic index/uuid base labels.

Closes #177.

• Rename uuid/index → npu_uuid/npu_index in all NPU exporters under src/api/metrics/npu/ and in the NPU-specific mock templates under src/mock/templates/.
• Leave the GPU-family base labels (gpu_uuid/gpu_index) on metrics that are shared across GPU and NPU devices (all_smi_gpu_*, all_smi_ane_*) since those are emitted by the unified exporter in src/api/metrics/gpu.rs.
• Extend the remote parser to fall back through gpu_uuid → npu_uuid → uuid and gpu_index → npu_index → index (with a matching gpu → npu device-name fallback), so mixed-version clusters continue to parse. Two new round-trip tests lock the aliasing and precedence order in.
• Update the common::add_basic_gpu_metrics mock helper to emit gpu_uuid/gpu_index (closing a gap left by PR fix: cross-PR review fixes for NVIDIA monitoring features (#172-#175) #176) so the mock stream now matches the real exporter for all_smi_gpu_* lines.
• Refresh API.md rows for all affected NPU metric families and extend the v0.21.0 entry in README.md to note the NPU rename.

Breaking change

The NPU base labels index and uuid in the Prometheus exposition format are renamed to npu_index and npu_uuid for all non-NVIDIA NPU-specific metrics (all_smi_npu_*, all_smi_tenstorrent_*, all_smi_rebellions_*, all_smi_furiosa_*, all_smi_gaudi_*, all_smi_tpu_*).

Existing Prometheus dashboards and alert rules that filter NPU metrics by index= or uuid= labels must be migrated to use npu_index= and npu_uuid= respectively.

Migration

- all_smi_tpu_utilization_percent{uuid="TPU-0", index="0"}
+ all_smi_tpu_utilization_percent{npu_uuid="TPU-0", npu_index="0"}

- all_smi_tenstorrent_asic_temperature_celsius{uuid="Tenstorrent-0", index="0"}
+ all_smi_tenstorrent_asic_temperature_celsius{npu_uuid="Tenstorrent-0", npu_index="0"}

The remote all-smi view parser accepts both the new and legacy label names for one release so that nodes running older builds can still be consumed by newer clients during a rolling upgrade.

Test plan

• cargo build passes
• cargo test — all 495 unit tests + 543 binary tests + integration suites pass
• New tests parser_accepts_npu_labels_for_uuid_and_index and parser_prefers_gpu_uuid_when_both_gpu_and_npu_labels_present cover the new aliasing and precedence
• cargo fmt clean
• Manual inspection confirms all_smi_gpu_*/all_smi_ane_* metrics continue to use gpu_uuid/gpu_index while all_smi_npu_*/vendor-specific metrics use npu_uuid/npu_index