[Lambda] Binary Invoke API

### Community Note

* Please vote on this issue by adding a 👍 [reaction](https://blog.github.com/2016-03-10-add-reactions-to-pull-requests-issues-and-comments/) to the original issue to help the community and maintainers prioritize this request
* Please do not leave "+1" or "me too" comments, they generate extra noise for issue followers and do not help prioritize the request
* If you are interested in working on this issue or have submitted a pull request, please leave a comment

**Tell us about your request**

It would be excellent to have a binary API for invoking Lambdas, and receiving responses.

**Which service(s) is this request for?**
This is a request for Lambda proper, but could benefit integrations with other services (AWS and 3rd party) as well.

**Tell us about the problem you're trying to solve. What are you trying to do, and why is it hard?**
Many use cases for Lambda involve processing binary data (network, images, data, etc.). The current REST API requires two full encoding/decoding passes:

- The calling code must encode the binary data to invoke the lambda.
- The Lambda function decodes the request to process it.
- The Lambda function encodes the response.
- The calling code decodes the response.

In addition to the encoding and decoding compute being a burden at scale, encoding inflates the size of payloads (hence reducing the effective payload size).

Removing the encoding would reduce complexity in customer code, reduce compute consumption and increase the effective request and response payload size.

**Are you currently working around this issue?**
At Proxylity we encode each UDP packet arriving at our service before passing to Lambda. An arriving packet is serialized to JSON with the binary data being base64 encoded before sending to Lambda. Our customers then write (or use library) code that handles the de/reserialization. Our code then decodes the JSON response and sends a the binary data in a UDP response. The encoding alone can account for as much as a third of the CPU time overall for simple handlers.

**Additional context**
I understand this ask may sound daunting, but could it perhaps be compatible with the current runtime API as an underlay (pre deserialization/serialization)?

**Attachments**
None at this time.