[Misc] Various cleanups for MM input processing

[DarkLight1337]

Purpose

• Add vllm.utils.serial_utils.tensor2base64 convenience function
• Fix some type annotations for MM embedding input parsing
• Fix Terratorch test not being run because of being incorrectly marked async and fix the test failure by allowing tokenizer=None in _preprocess_chat
• Remove unused tokenizer argument for MultiModalItemTracker classes, and by extension parse_chat_messages functions

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

Documentation preview: https://vllm--29970.org.readthedocs.build/en/29970/

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[gemini-code-assist]

Code Review

This pull request introduces several valuable cleanups for multimodal inputs. The addition of the tensor2base64 utility function is a great improvement, simplifying code that serializes tensors to base64 and reducing duplication across documentation, examples, and tests. The removal of the unused tokenizer argument from parse_chat_messages and parse_chat_messages_futures cleans up the API nicely. I also appreciate the fix for the Terratorch test and the addition of missing type annotations.

I have one minor suggestion to improve the new utility function further. Overall, this is a solid set of changes that improves code quality and maintainability.

[DarkLight1337]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces several cleanups and fixes related to multi-modal inputs. A new convenience function tensor2base64 is added to simplify tensor serialization, and it's consistently used across documentation, examples, and tests. A significant refactoring removes the unused tokenizer argument from MultiModalItemTracker and related functions, which cleans up the API. The PR also includes fixes for incorrect type annotations in multi-modal data parsers and corrects a test that was improperly marked as asynchronous. Overall, these changes improve code quality, consistency, and correctness. The implementation looks solid, and I have no further suggestions.

[lgeiger]

Should we also use pybase64 here as introduced in #19984, or do you think it's not worth for this case.

[DarkLight1337]

Let's do that in a separate PR