fix(privacy): gate cloud usage reporting to OpenWhispr-cloud transcriptions

[xpipko]

The problem

The streaming completion path reported usage to the OpenWhispr API for every
provider, and that call carries sendLogs — the transcript text itself.

For BYOK providers, custom endpoints, and self-hosted servers the user has
deliberately routed audio away from OpenWhispr's cloud. Shipping their
transcripts back defeats the point of choosing those providers.

The fix

Providers declare in the streaming registry whether their audio travels through
OpenWhispr Cloud, and the session records the answer at connect time from the
provider it selected and the mode it sent:

this.streamingSessionMetered =
  !!provider.cloudMetered && sessionOptions.mode === "openwhispr";

Completion reads that recorded decision instead of re-deriving routing from
settings, so a mid-dictation settings change can no longer decide where the
transcript goes.

mode is read back off the object buildStreamingSessionOptions() built rather
than re-derived, so the usage decision and the connect payload cannot disagree —
and if that builder ever changes how it resolves the mode, this follows
automatically.

The flag is opt-in. A backend added later stays silent until someone states
otherwise, so the failure mode of an oversight is a missing usage row rather
than a leaked transcript.

Provider	Reports usage	Why
deepgram	yes	OpenWhispr Cloud is the upstream
assemblyai	yes	OpenWhispr Cloud is the upstream
openai-realtime	only in openwhispr mode	runs both as Cloud's own upstream and, in BYOK, against the user's key
corti	no	streams over Corti's own WSS on the user's credentials
tinfoil-realtime	no	streams against the user's own Tinfoil key
local / self-hosted	n/a	never open a streaming session

Tests

test/helpers/audioManagerUsageGate.test.js drives production code rather than
restating it: each case sets a settings state and runs the real
resolveStreamingProviderName(), the real STREAMING_PROVIDERS registry via
getStreamingProvider(), and the real buildStreamingSessionOptions(), then
asserts which combinations may report usage. Six cases cover all five streaming
providers, including openai-realtime in both modes.

Confirmed it catches the regression it exists for: dropping cloudMetered from
openai-realtime fails the OpenWhispr-managed case.

It pins the gate's inputs, not the gate expression or its single consumer — both
sit inside startStreamingRecording past the mic and worklet setup, where there
is no seam to call. So it catches a backend added without a decision, or a BYOK
provider flipped to metered; it would not catch the guard itself being deleted.

npm run lint clean, full suite green (2089 tests).