Bedrock provider improvements

[badlogic]

Follow-up improvements for the Amazon Bedrock provider (#494):

Tests for Non-Claude Models

The provider currently only has comprehensive tests for Claude models. Add tests for:

• Amazon Nova (Micro, Lite, Pro, Premier)
• Meta Llama models
• Mistral models
• AI21 Jamba models
• Cohere Command R models

Thinking Support for Non-Claude Models

Some non-Claude models support reasoning (DeepSeek R1, Nova Premier, Qwen3) but use different mechanisms:

• DeepSeek R1: Uses <think> tags in prompt/response
• Others: May have vendor-specific parameters

Investigate and implement where feasible.