Support streaming model import to avoid OOM for large (multi-GB) files

[sanchitmonga22]

Problem

ModelManager.importModel() and the model loading pipeline read entire files into memory via file.arrayBuffer(). For 2-8 GB LLM models, this spikes JS heap memory and can cause OOM crashes on constrained devices.

Current State

• importModel() calls new Uint8Array(await file.arrayBuffer()) — full file in memory
• ModelLoadContext.data: Uint8Array forces loaders to receive the full file
• Double-buffering: JS heap copy + WASM linear memory copy

Proposed Solution

1. Add streaming interface to ModelLoadContext: dataStream?: ReadableStream<Uint8Array>
2. Update importModel() to use file.stream() and pipe chunks to storage
3. When LocalFileStorage is active, avoid copy entirely by passing the File handle
4. Update backend loaders to support chunked writes to their WASM FS

Impact

• High for users downloading large LLMs (2+ GB)
• Medium complexity — requires interface changes across core + backends

From PR #370 review comments (greptile + coderabbit).

[Harin22]

I’m interested in helping with this.. I understand the current issue is that model import loads the full file into memory, which can cause OOM for large models. As I’m still new here, what small task would be most useful for me to take on? Please let me know..