feat(backends): add Whisper-Medusa transcription

[localai-org-maint-bot]

What does this PR do?

Adds a dedicated Python gRPC backend for aiola Whisper-Medusa checkpoints.

• loads local paths or Hugging Face model identifiers through the standard model resolver
• normalizes input to mono 16 kHz audio and enforces upstream’s 30-second clip limit
• exposes language and generation regulation options
• adds Linux amd64 CPU and NVIDIA CUDA 12 backend images
• registers the backend in the backend gallery and Make targets
• documents configuration, supported checkpoints, and upstream limitations

The upstream project is archived and pins PyTorch 2.2.2. This PR intentionally does not advertise macOS, ROCm, Intel GPU, or Jetson images because upstream has not published compatibility for those targets.

Closes #3127

Testing

• python3 -m unittest discover -s backend/python/whisper-medusa -p "test_unit.py"
• python3 -m py_compile backend/python/whisper-medusa/backend.py backend/python/whisper-medusa/test_unit.py
• git diff --check
• YAML parse of .github/backend-matrix.yml and backend/index.yaml with gopkg.in/yaml.v3
• make -n docker-build-whisper-medusa BUILD_TYPE= BASE_IMAGE=ubuntu:24.04

The focused Go gallery suite could not run because generated protobuf stubs are absent in a fresh worktree and the pinned protoc download was rejected by the environment proxy with HTTP 403. A full backend image build was not run because it downloads the multi-gigabyte Whisper-Medusa model/dependency stack.

Checklist

• Tests added for new backend helpers
• Backend build matrix and gallery metadata updated
• User-facing documentation added
• No coverage baseline or tolerance changes