Remote detection backend improvements: Roboflow fix, resize, CPU fallback, coralapi

[nathan-v]

Description

Four related improvements to the remote detection backend layer:

• Roboflow request-format fix: Inference Server 1.x requires model_id and image in the JSON body. The old query-param format gets HTTP 422, which surfaced as clips "successfully" processing with zero detections.
• Client-side downscaling (roboflow.resize_max_px, coralapi.resize_max_px, opt-in): the server resizes to model input anyway, so smaller uploads cut per-frame latency 3-4x at identical detection quality. Predictions are rescaled to original frame coordinates so calibration is unaffected.
• CPU fallback (detection_fallback: local, opt-in): remote detectors now raise DetectorUnavailableError on transport failures instead of silently returning empty detections. A FallbackDetector wrapper switches to local CPU inference (immediately if the backend never succeeded this job, after 3 consecutive failures otherwise; sticky per job; local model loaded lazily). ProcessingResult gains detection_backend/detection_model/detection_fallback fields recording what actually ran, and these are emitted as backend/model/fallback tags on InfluxDB metrics.
• New coralapi backend (detection_backend: coralapi) for Coral Edge TPU inference via coralapi: multipart /v1/vision/detect, normalized-box scaling, label mapping with a COCO-90 index fallback for label-less models.

Related Issue

N/A

Motivation and Context

The Roboflow format bug and the silent-empty-detection failure mode were both found in production. Resize was benchmarked at 3-4x per-frame speedup at every model size. Fallback protects a ~350 clip/day pipeline from remote host outages (laptop sleep and a desktop outage both occurred during development). The metric tags enable per-model dashboards and fallback-rate alerting.

How Has This Been Tested?

• 166 unit tests pass (pytest); ruff check and format clean
• New tests: Roboflow wire format, resize/box-rescaling for both backends, fallback policy matrix (immediate/threshold/sticky/lazy/no-fallback), config validation, metric tag emission, coralapi wire format and COCO-90 index fallback
• Live: deployed to a production cluster; Roboflow backend verified against GPU inference hosts; fallback exercised end-to-end against a dead backend (real detections via CPU, warning surfaced, tags correct); coralapi backend benchmarked against a Coral USB Accelerator across 8 models, including a byte-identical regression between label-path and index-fallback-path class mapping

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• My code follows the code style of this project (ruff check and ruff format pass)
• My change requires a change to the documentation
• I have updated the documentation accordingly
• I have read the CONTRIBUTING document
• I have added tests to cover my changes
• All new and existing tests passed