fix(metrics): emit gpu_usage_percent as float for Cloud Monitoring (FILTER-585)

[shingonoide]

📋 What does this PR do?

Casts the gpu_usage_percent system-health metric to float so it is always emitted as a floating-point value.

- self.gpu_usage_percent = gpu.get('gpu0', 0.0)
+ self.gpu_usage_percent = float(gpu.get('gpu0', 0.0))

🔍 Why is this needed?

NVML reports GPU utilization as an integer, so gpu_usage_percent was emitted as an integer whenever a GPU was active (the float default 0.0 only applies when no GPU is present). The Google Cloud Monitoring descriptor for this gauge is type-locked to DOUBLE, so every active-GPU point was rejected by the backend with:

value type for metric must be DOUBLE, but is INT64

and silently dropped (a partial, single-point rejection per export batch). The result is silent gaps in GPU-utilization data. Forcing the value to float makes the emitted points match the existing DOUBLE descriptor, restoring ingestion with no descriptor migration.

The change is intentionally narrow: the per-device <filter>_gpuN counters and the legitimately-integer gpu_accessible / camera_connected metrics are left untouched, since their descriptors are integer-typed and casting them would cause the opposite rejection.

🧪 How was it tested?

• New unit test tests/test_gpu_usage_percent_float.py drives the GPU poll path with a mocked integer NVML reading and asserts gpu_usage_percent is a float equal to 55.0. It fails on the previous (un-cast) code and passes with this change.
• Existing GPU / telemetry / timing suites remain green (tests/test_gpu.py, tests/test_telemetry.py, tests/test_timing_metrics.py).

🔗 Related Issues

FILTER-585

✅ Checklist

• I have read and agreed to the terms of the LICENSE
• I have read the CONTRIBUTING guide
• I have followed the coding style
• I have signed all commits in compliance with the DCO (git commit -s)
• I have added or updated tests as needed
• I have added or updated documentation as needed

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[leandrobmarinho]

Review

Summary: Casts gpu_usage_percent to float in Metrics.gpu_thread_func so an active GPU's integer gpu_util (gpu.py:71, gpu_util: int) no longer gets rejected by the DOUBLE-locked openfilter_gpu_usage_percent Cloud Monitoring descriptor (value type ... must be DOUBLE, but is INT64).

Verified the fix is correct and minimal: gpu['gpu0'] derives from the int gpu_util when a GPU is present, so the old code emitted an int; the cast forces DOUBLE. Scope is right — the per-device gpuN series and the legitimately INT64 descriptors (gpu_accessible, camera_connected) are untouched, and the downstream observability/client.py handles the value generically. No contract break, no reference/mutation concern.

FYI (non-blocking): test_gpu_usage_percent_is_float_when_gpu0_absent doesn't actually guard the regression — the no-gpu0 default 0.0 was already a float on both old and new code, so it passes either way; the gpu0-present test is the one that exercises the fix. No change needed to merge.

[lucasmundim]

Reviewed: the float() cast is correct and narrowly scoped to the gpu_usage_percent summary field that maps to Cloud Monitoring's DOUBLE descriptor (openfilter_gpu_usage_percent). Per-device gpuN series and the legitimately-INT64 flags are correctly left untouched. VERSION bump, RELEASE.md format, and the new test all check out. LGTM.