fix(memory): add dedicated SGLang and vLLM thinking transports

[LeslieWylie]

Summary

Add dedicated sglang and vllm memory LLM providers for the Qwen chat_template_kwargs.enable_thinking transport.

The generic openai_compatible provider keeps its existing top-level request shape, so cloud-provider calls are unaffected.

Changes

• accept sglang and vllm as explicit LLM provider values
• route both providers through the existing OpenAI-compatible chat-completions client
• move Qwen enable_thinking into chat_template_kwargs only for those dedicated providers
• preserve caller-supplied chat_template_kwargs, including an explicit enable_thinking override
• keep top-level enable_thinking behavior unchanged for DashScope and generic OpenAI-compatible endpoints

Validation

• vitest run Memory/tests/llm-thinking.test.ts — 21 tests passed
• coverage includes SGLang, vLLM, disabled/enabled thinking, preserved template kwargs, and generic-provider isolation
• live SGLang replay previously confirmed that top-level enable_thinking=false was ignored while chat_template_kwargs.enable_thinking=false produced zero reasoning tokens and valid JSON

[Wang-Daoji]

Thanks for working on this! I suggest adding chat_template_kwargs later as part of dedicated SGLang/vLLM providers. Adding it to the generic OpenAI-compatible path now could affect existing cloud-provider calls and introduce unnecessary risk.

[LeslieWylie]

Thanks — I reworked the PR around that boundary. chat_template_kwargs is now used only by explicit sglang and vllm providers; the generic openai_compatible path keeps its existing top-level request shape. I also added coverage that pins the generic-provider isolation.

[LeslieWylie]

Superseded by clean-history replacement #348. An amend from a shallow checkout accidentally removed this branch's base ancestry, so GitHub closed the PR and would not reopen it after the ancestry was repaired. The replacement keeps the maintainer-requested provider isolation and the passing 21-test validation.