perf: cannon-rotated WASM SIMD kernels (14KB→22KB)

perf: raise token limits to 256 (both PyTorch and Aether)

feat: Aether v2 with RoPE fix -- PyTorch vs Aether side by side

fix: remove missing WASM flashAttention + stream results independently

fix: handle WASM OOM for LM head + show error details

fix: use Q8_0 GGUF + type-aware dequantization

perf: add WASM SIMD kernels + use Q4_K_M for faster inference

feat: PyTorch vs Aether side-by-side inference