fix(models): give QwenVLModel a timeout, wired to DashScope's request_timeout

[AmirF194]

QwenVLModel is the one model class with no timeout knob. Every DashScope call was pinned to the SDK's own 300s default, and passing timeout= to the constructor raised a TypeError.

Worth flagging: the fix isn't just "add a timeout kwarg like OpenAIChatModel does." DashScope's MultiModalConversation.call() accepts a timeout= kwarg but drops it into the request body as an unused parameter; the actual socket timeout only applies through request_timeout=. I checked this by patching HttpRequest.__init__ and confirming timeout=5.0 never reaches it, while request_timeout=5.0 does. So the constructor still takes timeout for parity with the sibling classes, but generate() forwards it as request_timeout internally.

I didn't add max_retries alongside it. DashScope's HTTP layer for this call has no retry mechanism at all, so a max_retries param would be exactly the kind of silent no-op this PR is fixing.

Added tests/models/test_qwen_vl_model.py (didn't exist yet) covering timeout set, unset, and explicit None, each asserting request_timeout (never timeout) is what reaches the mocked call. Ran the full tests/models/ suite on python 3.11, 57 passed / 3 skipped (integration tests need live keys), black and isort clean.