Fix conv_template: use full object instead of string name (fixes web-llm compat)

Revert tensor-cache.json dataPath to root (shards now at repo root)

Upload param shards to repo root for WebLLM direct access

Fix tensor-cache.json: prefix dataPath with params/ for correct shard URLs

Add tensor-cache.json to repo root for WebLLM

Add MLC/WebGPU compiled model for WebLLM

initial commit