Whisper ASR stack v0.1.0

The first public release of the Whisper ASR stack: self-hosted speech recognition with Speaches and faster-whisper.

• Pinned upstream container and persistent model storage.
• Source-hash-checked fix for the explicit model unload deadlock.
• Alias resolution on unload, including whisper-1.
• Three CPU regression tests covering callback completion, active-model protection and the actual DELETE route.
• Shared-GPU integration with the inference SDK.

The lifecycle fixes were checked through a synthetic GPU handoff, including transcription after model reload. These checks do not establish WER or throughput.

Build the image from the tagged source using the quickstart. See the lifecycle explanation for the bugs and regression strategy. Model weights are not included in this release.