Add OFFLINE_MODE for air-gapped deployments

[alexnorell]

Summary

• Adds OFFLINE_MODE env var that blocks all outbound Roboflow API traffic
• Models and workflows load from local cache only -- no retries, no timeouts
• Composes with existing feature flags (METRICS_ENABLED, DISABLE_VERSION_CHECK, ACTIVE_LEARNING_ENABLED, SINGLE_TENANT_WORKFLOW_CACHE) rather than adding a parallel guard system
• Guards at chokepoints: _get_from_url() and post_to_roboflow_api() for sync requests, individual POST functions for fire-and-forget operations, AutoModel.from_pretrained() and get_model_from_provider() for inference_models model loading
• Workflow spec falls back to file cache via the existing except ConnectionError handler naturally
• Bumps inference-models to 0.27.3

How it works

Flag composition (env.py): OFFLINE_MODE=True auto-sets METRICS_ENABLED=False, DISABLE_VERSION_CHECK=True, ACTIVE_LEARNING_ENABLED=False, SINGLE_TENANT_WORKFLOW_CACHE=True.

Chokepoint guards (roboflow_api.py): _get_from_url() raises ConnectionError and post_to_roboflow_api() raises RoboflowAPIConnectionError, which flow through existing error handlers (e.g. get_workflow_specification falls back to file cache without any OFFLINE_MODE-specific code). Fire-and-forget endpoints (image upload, annotation, monitoring) return empty/no-op.

Usage tracking: UsageCollector._offload_to_api() and PlanDetails fetch/WebRTC lookups short-circuit silently.

inference_models: get_model_from_provider() raises ModelRetrievalError, download_files_to_directory() raises RuntimeError. AutoModel.from_pretrained() has a dedicated attempt_loading_model_from_offline_cache() fallback that scans {INFERENCE_HOME}/models-cache/ for cached packages. Also handles RetryError (API unreachable) with the same offline cache fallback even when OFFLINE_MODE is not set.

Customer usage

# docker-compose.yml -- phase 1: cache models with network
services:
  inference:
    environment:
      ROBOFLOW_API_KEY: "your-key"
      SINGLE_TENANT_WORKFLOW_CACHE: "True"

# docker-compose.yml -- phase 2: run offline
services:
  inference:
    environment:
      OFFLINE_MODE: "True"
    # no ROBOFLOW_API_KEY needed
    # no network needed

Test plan

• 138 roboflow_api unit tests pass
• 836 inference_models unit tests pass (35 new)
• 935 inference core unit tests pass
• black, isort, flake8 clean
• End-to-end container test: private model lildc-hardhat/4 loaded from cache with --network none, no API key
• Workflow spec loaded from file cache offline
• Uncached model/workflow fails immediately