[codex] fix Gemma 4 thinking generation prompt

[romgenie]

What changed

Fixes the Gemma 4 generation prompt in both shipped Gemma 4 templates so thinking mode opens the thought channel for generation instead of closing an empty thought block. Non-thinking mode now leaves the model turn open without injecting <|channel>thought\n<channel|>.

Why

The previous guard was inverted: it emitted a closed empty thought channel when enable_thinking was false. In practice this prevented Gemma 4 from producing visible reasoning when thinking was enabled and leaked the wrong control tokens into non-thinking prompts.

Validation

• Built test-chat in an Ubuntu 24.04 Podman container and ran ./build-test/bin/test-chat --template google-gemma-4.
• Ran the full ./build-test/bin/test-chat suite.
• Built the Vulkan server image with .devops/vulkan.Dockerfile using UBUNTU_VERSION=24.04 as localhost/llama.cpp:server-vulkan-latest-gemma4-test.
• Smoke-ran the image with llama-server --version, which loaded the Vulkan backend and reported version: 8981 (d77599234).