AWS: Refactor provider to use Bedrock Converse API

[Copilot]

The AWS provider had three separate model-family-specific implementations (Anthropic/Claude, Llama/Mistral, Cohere), each with their own request/response format. The Converse API provides a unified interface across all Bedrock chat models.

Changes

Unified chat execution (aws.ts)

• Replaced executeAnthropicModel, executeLlamaStyleModel, executeCohereModel (and their streaming counterparts + all helper methods) with a single executeWithConverse() / streamWithConverse() pair
• Message/system/tool conversion now happens through a single convertRequestToConverse() pipeline rather than three divergent paths
• Hook types updated from InvokeModelCommand | InvokeModelWithResponseStreamCommand → ConverseCommand | ConverseStreamCommand
• Dropped anthropic-specific config block (version, beta, emptyMessage) — not applicable to the Converse API
• Fixed broken imports (isModelInfo from wrong path, unused http module, Anthropic SDK types)
• Streaming now handles reasoningContent blocks natively via Converse API events

// Before: routing to family-specific implementations
if (family === 'anthropic') return self.executeAnthropicModel(...)
else if (family === 'meta' || family === 'mistral') return self.executeLlamaStyleModel(...)
else if (family === 'cohere') return self.executeCohereModel(...)

// After: single Converse API path for all chat models
return self.executeWithConverse(modelInput, request, ctx)

Capability detection (common.ts)

• Expanded tools capability to Llama 3.1+, Mistral Large, and Cohere Command R — models that support tool use via the Converse API but were previously excluded

Removed dependency

• @anthropic-ai/sdk devDependency dropped; Anthropic-specific SDK types are no longer needed

Image generation (Stability AI) and embeddings (Amazon Titan) continue to use InvokeModelCommand — they are not part of the Converse API surface.