Bug: AWS-Anthropic custom provider sends deprecated Anthropic request fields

## Bug: AWS-Anthropic custom provider sends deprecated Anthropic request fields

### Environment
- GitHub Copilot App: 1.0.80
- OS: macOS
- Provider type: AWS-Anthropic
- Wire API: Anthropic Messages
- Model display name: claude-opus-4-8 / claude-opus-5
- Wire model sent to provider: anthropic.claude-opus-4-8 / anthropic.claude-opus-5
- Custom base URL: https:///anthropic
- Header: anthropic-version: 2023-06-01

### Steps to reproduce
1. Configure an AWS-Anthropic custom provider with an Anthropic Messages-compatible endpoint.
2. Configure Claude Opus 4.8 or Opus 5 using an `anthropic.` wire model ID.
3. Send a simple prompt such as “你好”.

### Actual result
The GitHub Copilot App sends legacy request fields. The upstream rejects them with:

- `400 temperature is deprecated for this model`
- `400 "thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.`

### Expected result
For current Claude models, the AWS-Anthropic provider should omit deprecated sampling fields (`temperature`, `top_p`, `top_k`) and use the current thinking protocol:
`thinking.type: "adaptive"` with `output_config.effort`, or provide a per-model capability/protocol setting.

### Impact
The Custom endpoint provider cannot be used as a workaround because it only offers OpenAI-compatible Completions/Responses wire APIs, not Anthropic Messages. This prevents use of current Claude Bedrock models through a custom Anthropic endpoint without external request rewriting.

### Evidence
Attached screenshots show:
- the AWS-Anthropic provider and model configuration;
- real Copilot request failures;
- request IDs from the failed requests.

---
| Field | Value |
| --- | --- |
| App version | 1.1.14 |
| OS | macOS 15.7.5 |
| Theme | GitHub |
| Path | /chat |
| Tenure | Day 6 (Week 1) |

未指定文件或测试。首先跟踪 AWS-Anthropic 自定义 provider 的 Anthropic Messages 请求路径及其模型能力处理，然后使用列出的 Claude wire model IDs 重现该请求。完成标准是当前模型不再接收已弃用的 sampling 字段，并使用预期的 adaptive thinking protocol，同时覆盖报告中的 configuration。

由索引模型根据 Issue 内容生成。