fix(gemini): forward stream chunks that carry only UsageMetadata

[kenijkawada]

Summary

Gemini 3 family models (e.g. gemini-3-flash-preview,
gemini-3.1-flash-lite-preview) emit usageMetadata on stream chunks that
contain no text and no function calls. The current gemini.StreamAdapter
drops those chunks before they reach the runtime, so all token counts and
costs come back as zero. Gemini 2.5 happens to fit its usageMetadata into
the same chunk as the final text, which is why it has worked so far.

Root cause

Two places in pkg/model/provider/gemini/adapter.go:

1. NewStreamAdapter forwarded only chunks with hasText || hasFuncs.
   Usage-only chunks were silently discarded before reaching the channel.
2. Recv() extracted UsageMetadata only in the else if res.resp != nil
   branch, so even when the synthesised "done" event carried the last
   response, the usage was not surfaced.

Fix

• Forward chunks that have UsageMetadata even when text and tool calls are
  absent (hasUsage check added to the OR).
• Move Usage extraction out of the else if branch so the done event
  also exposes the usage of the last response.

Reproduction

gemini-3-flash-preview Vertex AI response contains usageMetadata (verified
with a direct curl against :generateContent and :streamGenerateContent),
but docker agent run --exec --yolo records usage = {input:0, output:0}
in session_items.message_json and sessions.cost = 0 in the SQLite store.

After the fix the same run records non-zero tokens and the cost computed by
the runtime against the models.dev catalog.

model	before	after
gemini-3-flash-preview	usage = all zero, cost = 0	usage populated, cost > 0
gemini-2.5-flash	already worked	still works (regression test)

Tests

pkg/model/provider/gemini/adapter_test.go gains three sub-tests under
TestStreamAdapter_GeminiUsageMetadata:

• forwards chunks containing only UsageMetadata — directly guards the
  dropped-chunk regression.
• done event carries usage from last response — guards the second fix
  (extracting usage on the done branch).
• trackUsage=false suppresses usage extraction — confirms the existing
  opt-out still works.

The existing TestStreamAdapter_FunctionCalls continues to pass.