Re-export with coreai-torch 0.4.1 (fixes OS 27 beta 2+ load failure); gated token-for-token vs fp32 oracle

Add b2 (beta3-loadable) bundle under gpu-pipelined-b2/qwen3_5_0_8b_decode_int8hu_block32_sym/ (b1 retained)

Upload folder using huggingface_hub

qwen3.5-0.8B int8lin decode-only loop-free bundle (pipelined engine): Mac 204 tok/s, iPhone 50.3-51.5 tok/s

ios-gpu: add qwen3_5_0_8b_ios_hc_prefill_q16_b2048_int8.aimodel

ios-gpu: add qwen3_5_0_8b_ios_hc0_int8v3.aimodel

macOS GPU best: dynamic int8 (58.5 tok/s release)

iOS ANE best: dynamic int8 (14.7 tok/s)

iOS GPU best: fp16 static ctx-2048 monolith (27.7 tok/s)

static ctx-2048 monolith (iPhone GPU 27.7 tok/s, release config)

dynamic int8 bundle (iPhone GPU 12.5 / ANE 14.7 / Mac 58.5 tok/s)

initial commit