Add Qwen3 VL model support

[amirgholipour]

What changed

• Add Qwen model-family detection to the shared Bedrock client.
• Convert text and image messages to the Bedrock Converse request format.
• Normalize Qwen text, usage, and stop-reason responses into the specialist response shape.
• Allow the orchestrating runtime and image enhancer to select direct model IDs through deployment environment variables.
• Grant direct regional invocation permissions for qwen.qwen3-vl-235b-a22b.
• Make application inference profiles optional for direct-model deployments.
• Add Qwen3 VL configuration guidance and limitations to the main README and deployment documentation.

Why this change is proposed

The model abstraction currently recognizes Claude and Nova families, while unknown model families fail before invocation. Multimodal Qwen models are available through Bedrock Converse and can process the same text-and-image specialist messages, but they require different request construction and do not use Claude thinking fields or application inference profiles.

This change extends the existing abstraction instead of adding a separate processing path. Existing profile-backed configurations remain the default, while Qwen can be selected explicitly per deployment and per specialist.

Benefits

• Adds another multimodal model option for document and image analysis.
• Preserves existing model defaults and specialist configuration structure.
• Supports direct regional invocation where application inference profiles are unnecessary or unavailable.
• Retains usage metadata returned by Bedrock for observability.
• Documents the exact model ID, configuration examples, deployment variables, and thinking-field limitation.

Validation

• Deployed the agent and image specialists with qwen.qwen3-vl-235b-a22b in an isolated BADGERS environment.
• Processed direct image input through classifier, enhancer, full-text, elements, charts, and correlation specialists across completed validation sessions.
• Verified Qwen response normalization with text and token-usage metadata.
• Completed an offline CDK synthesis with inference profiles omitted and Qwen selected for the agent and enhancer.
• Verified all changed Python modules compile.
• Reviewed the branch against the current upstream main.

Authored by Saeed Kasmani (amirgholipour).