Increase AI Gateway embeddings HTTP timeout above 20s

Description

The Gitlab::Llm::Embeddings::Client calls the AI Gateway embeddings endpoint via Gitlab::HTTP.post without setting a read_timeout, so it uses GitLab's default of 20 seconds. On July 29, 2026, Vertex AI (the embeddings backend) experienced degraded performance with response times of 19–22 seconds, exceeding this timeout. This caused the Semantic Code Search embeddings pipeline to fail with Net::ReadTimeout errors, which built up large retry and dead-letter queues. We now explicitly set read_timeout: 30.seconds on that HTTP call to provide headroom for slow-but-successful responses. A spec verifies that the timeout option is passed to Gitlab::HTTP.post.

Related to #607802 (closed).