fix(go/plugins/googlegenai): do not resend the cached prefix inline

[a2105z]

Summary

• Context caching stored messages[0..endIndex] on the CachedContent resource but toGeminiContents still sent those same messages in the generate request.
• The JS plugin slices the cached span off the request; Go did not, so the prefix was billed as cached tokens and as full-price inline input, and the model saw it twice.
• handleCache now returns the cache boundary. toGeminiContents skips messages at or before that index (-1 means no cache).

Fixes #6137

Test plan

• Unit test reproduces the bug: with a cache marker on message 0, sending the whole request still includes the prefix
• Same test asserts the generate contents after the cache boundary are only the follow-up
• Existing TestToGeminiContents still converts the full request when cachedThrough is -1
• go test ./go/plugins/googlegenai/
• CLA check

[gemini-code-assist]

Code Review

This pull request updates the README.md for the compat-oai plugin to reflect the new import path for the OpenAI plugin (@genkit-ai/compat-oai/openai) and the use of openAI.model() and openAI.embedder() helpers. It also updates the custom models section to demonstrate using openAICompatible with openrouter and compatOaiModelRef. A review comment suggests removing an unused import z from 'genkit' in the custom models code example.

[gemini-code-assist]

The z object is imported from 'genkit' but is not used anywhere in this code example. We can simplify the import by removing it.

Suggested change

	import { genkit, z } from 'genkit';
	import { genkit } from 'genkit';

[a2105z]

Hi @genkit-ai/genkit-googlers @pavelgj @apascal07 @huangjeff5 @ifielker @MichaelDoyle @yesudeep @ssbushi @shrutip90 @cabljac @adesinah @xcapaldi — this is the fix for #6137.

Context caching was writing the prefix into CachedContent and still sending it inline. toGeminiContents now skips the cached span, same as JS.

Could someone take a look when you have a moment? Happy to adjust anything. Thank you!