Add model selector to compare speed of smaller Moonshine models

[monneyboi]

Motivation

Dictation currently always runs the medium streaming Moonshine model (ModelSpec.stt("en", JNI.MOONSHINE_MODEL_ARCH_MEDIUM_STREAMING, false) in MoonshineModel.kt). We want to measure whether smaller models are fast enough / faster on target devices, and what accuracy we trade away. A model selector would let us switch models and compare.

Current behavior

• MoonshineModel.spec() hardcodes MOONSHINE_MODEL_ARCH_MEDIUM_STREAMING.
• The SDK also exposes MOONSHINE_MODEL_ARCH_TINY, MOONSHINE_MODEL_ARCH_TINY_STREAMING, MOONSHINE_MODEL_ARCH_BASE, MOONSHINE_MODEL_ARCH_BASE_STREAMING, MOONSHINE_MODEL_ARCH_SMALL_STREAMING.
• Because spec() determines the cache directory, only the medium model is ever downloaded or used.

Desired behavior

• A way to pick which Moonshine model the IME uses (at minimum: tiny/base/small/medium streaming variants), so we can A/B latency and accuracy of dictation on real devices.
• Download flow in setup must cover the selected model (per-model cache directories already fall out of ModelCache.directoryFor(context, spec(), null)).
• Keep the Moonshine contract intact: spec → configure → load() → start(), blocking calls serialized on the IME's worker executor; if the model can change while the IME is running, stop/reload must respect the existing STOPPING barrier.

Open questions

• UI surface: this MVP deliberately excludes settings screens. Should the selector be (a) a debug-only affordance (e.g. BuildConfig.DEBUG-gated), (b) a simple spinner on the setup screen, or (c) driven by a build config field for per-build testing?
• Should selection persist across IME restarts, and how (no transcript/settings persistence is stored today)?
• What metrics do we actually want to capture while testing — just subjective latency, or also measured time-to-first-partial / time-to-final?

Acceptance criteria

• Model choice is configurable in-app or per-build (per decision above)
• Setup downloads the selected model; IME uses the same one (no medium/other mismatch, since a missing model means a blocking download/failure at dictation time)
• Switching models mid-session behaves sanely (reload or restart, no leaked transcriber)
• Default remains the current medium streaming model unless changed