fix: support http_client factory in OpenAIModel client_args

[giulio-leone]

Problem

When http_client is passed as an httpx.AsyncClient instance in client_args, it gets closed after the first async with block. Subsequent requests fail because the same closed client instance is reused.

Users cannot pass a custom HTTP client to the OpenAI provider without it breaking on the second invocation.

Closes #1036

Solution

Support http_client as either:

• Instance (existing behavior): passed through unchanged — the httpx.AsyncClient is used directly
• Callable factory (new): invoked on every request to produce a fresh client

The distinction uses hasattr(http_client, 'send') — real HTTP clients have a .send() method, while factory functions/lambdas do not.

Usage

# Factory pattern — fresh client on every request (recommended)
model = OpenAIModel(
    model_id="gpt-4o",
    client_args={
        "api_key": "your-key",
        "http_client": lambda: httpx.AsyncClient(timeout=30),
    },
)

# Instance pattern — still works, caller manages lifecycle
model = OpenAIModel(
    model_id="gpt-4o",
    client_args={
        "api_key": "your-key",
        "http_client": my_custom_client,
    },
)

Changes

• src/strands/models/openai.py: Updated _get_client() to resolve factory callables; updated docstring
• tests/strands/models/test_openai.py: 3 new tests (factory invocation, instance passthrough, no mutation)

Testing

• All 1869 tests pass
• New tests verify factory is called on each request, instances pass through unchanged, and original client_args dict is not mutated

[giulio-leone]

Friendly ping — adds support for http_client factory functions in OpenAIModel client_args, enabling custom transport/proxy configuration.