new: [dependencies] split ML into optional extras, reduce default image from 12GB to 593MB

[wolffhechel]

Summary

Split ML dependencies (torch, transformers) into an optional extras group. The default install provides all 14 API-backed tools without PyTorch. ML classification tools (classify_severity, classify_cwe) are available when installed with -E ml.

Default Docker image: 593 MB (down from 12.1 GB).

Changes

• pyproject.toml: ML dependencies moved to [tool.poetry.extras]
• server.py: graceful registration - skips ML tools when torch is unavailable
• severity.py: absorbs classify_cwe from cwe.py (both ML-dependent)
• cwe.py: retains only get_recent_vulnerabilities_by_cwe (API-backed, no ML)
• Dockerfile: Python 3.12, three-mode ML install via INSTALL_ML build arg
• README.md: documents optional ML install, Docker variants, nvidia-docker
• .dockerignore: added

Install paths

poetry install              # 14 API tools, no torch (~500 MB)
poetry install -E ml        # 14 API tools + 2 ML classifiers
docker build .              # 593 MB, API tools only
docker build --build-arg INSTALL_ML=cpu .   # ~1.6 GB, CPU-only torch
docker build --build-arg INSTALL_ML=gpu .   # ~5+ GB, CUDA torch

The gpu variant requires nvidia-container-toolkit and docker run --gpus all.

Testing

• Default install: 14 tools registered, no torch/transformers imported
• ML install: 16 tools registered, classifiers functional with CPU torch
• Docker default: 593 MB image, healthy, 14 tools
• Docker ML (cpu): CPU-only torch, no nvidia-*/triton packages