feat(aws): add reasoning_effort parameter to ChatBedrockConverse

[thromel]

Summary

• Add reasoning_effort parameter to ChatBedrockConverse for simplified extended thinking configuration
• Support for "low", "medium", "high" effort levels
• Automatic translation to provider-specific formats (Amazon Nova, OpenAI on Bedrock)

Why

This addresses langchain-ai/langchain-aws#788 where users requested a simpler interface for configuring reasoning effort instead of manually setting additional_model_request_fields.

Changes

1. Added reasoning_effort parameter with type Literal["low", "medium", "high"] | None
2. Added _configure_reasoning_effort validator that translates the parameter to:
   • Amazon Nova: {"reasoningConfig": {"type": "enabled", "maxReasoningEffort": value}}
   • OpenAI on Bedrock: {"reasoning_effort": value}
   • Other providers: Passes through with a warning
3. Added comprehensive unit tests for all scenarios

Example Usage

from langchain_aws import ChatBedrockConverse

# Amazon Nova with reasoning
llm = ChatBedrockConverse(
    model="us.amazon.nova-lite-v1:0",
    region_name="us-west-2",
    reasoning_effort="medium",
)

# OpenAI on Bedrock with reasoning
llm = ChatBedrockConverse(
    model="openai.gpt-4o-mini-2024-07-18-v1:0",
    region_name="us-west-2",
    reasoning_effort="high",
)

Test Plan

• Unit tests for Amazon Nova reasoning configuration
• Unit tests for OpenAI on Bedrock reasoning configuration
• Unit tests for merging with existing additional_model_request_fields
• Unit tests for precedence when fields conflict
• Unit tests for warning on unsupported providers
• All existing tests continue to pass (122 passed)

Areas for Careful Review

• The validator merges reasoning_effort config with existing additional_model_request_fields, giving precedence to user-provided fields
• Warning is emitted for providers that may not support this parameter (e.g., Anthropic Claude which uses thinking.budget_tokens instead)