Add weightlift — in-browser ML model load progress library

[wassgha]

Wires Rescript to published weightlift@^0.2.1 with a registry-first API and optional Transformers.js load plumbing.

Rescript

import { transformersModel, fallbackDevicePolicy } from "weightlift/transformers";

const asrModels = new ModelManager({
  models: {
    [id]: transformersModel({
      pipeline,
      task: "automatic-speech-recognition",
      modelId: id,
      dtype,
      cacheKey: env.cacheKey ?? "transformers-cache",
      onDevice: (d) => { asrDevice = d; },
    }),
  },
});

await asrModels.load(MODELS[choice].id);
// WebGPU loss → fallbackDevicePolicy.preferWasm(); asrModels.unloadAll();

No per-app device policy setup — transformersModel uses shared fallbackDevicePolicy by default. UI status strings stay local in the worker. No monorepo / workspaces.