fix(providers): emit explicit disabled thinking when required

[GodfatherPacino]

Summary

• add an emit_disabled_thinking provider compatibility option
• emit thinking: { "type": "disabled" } on the Anthropic wire path when the option is enabled
• preserve the existing omission behavior by default for providers where omission already disables thinking
• document the option for Anthropic-compatible gateways

Problem

Some Anthropic-compatible gateways, including GLM-5.2 deployments, enable thinking by default. The runtime state could be switched to disabled, but the Anthropic request projector omitted the thinking field for ThinkingConfig::Disabled. The gateway therefore continued returning thinking content.

GLM documents that disabling thinking requires an explicit thinking.type = disabled request. The Anthropic Messages schema also defines a disabled thinking configuration, while some provider/model combinations still rely on omission or do not support disabling. A compatibility flag keeps that distinction configuration-driven.

Verification

• cargo fmt --all
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p aion-config compat_test (28 passed)
• cargo test -p aion-providers projector_test (27 passed)
• full repository just push gate, including nextest and cargo-hakari verification

References

• https://docs.bigmodel.cn/cn/guide/capabilities/thinking-mode
• https://platform.claude.com/docs/en/api/messages/create