[Core] Rework dtype resolution

[DarkLight1337]

Key changes:

• float32 models will be downcasted to whatever dtype is prioritized by the platform (defined by the first item in current_platform.supported_dtypes). This means that models may be downcasted to bfloat16 instead of float16.
• Attempt to detect safetensors dtype from HF Hub repo if the dtype is not specified in the configuration file. This avoids accidentally casting float16 weights to bfloat16 if the dtype is not inside the config on HF Hub.
• Exclude some models (gemma3, plamo2, glm4) from being downcasted to float16 because of reported numerical stability issues.
• Clean up some related log messages

Related to #17123

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @DarkLight1337.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork