fix(provider): keep Gemini requests from ending with a model turn - #55382

Fixes #47034 (same root cause as #45359) on the V1 line.

Problem

Gemini rejects any request whose contents ends with a model turn:

400 INVALID_ARGUMENT: Requests ending with a model turn are not supported.

V1 produces that shape whenever it continues from a persisted assistant tail:

• max-steps priming appends MAX_STEPS_PROMPT as an assistant message
• unknown-finish and empty-output retries replay history that ends with the assistant message

Fix

Append a synthetic user turn ("Continue.") in ProviderTransform.message when the last message is an assistant turn, for the Gemini SDKs only (@ai-sdk/google, @ai-sdk/google-vertex):

• outbound, in-memory only — the synthetic turn is never persisted, so transcripts and subsequent requests are unchanged
• appending at the tail leaves the provider's implicit prefix cache intact
• other providers keep true assistant prefill
• mirrors the existing wire-sequence repair in normalizeMessages, which inserts a synthetic assistant "Done." when a tool turn is followed by a user turn

Testing

• bun test test/provider/transform.test.ts — 585 pass, 0 fail
• bun typecheck (packages/opencode) — no new errors vs pristine v1 tip (42 pre-existing, verified via stash comparison)