[Fix] offload .to(cuda) interception + configurable nd_tiling max_tiles and prefetch lookahead

[cennn]

Summary

Three bug fixes and one config knob exposed during disaggregated multi-model inference on 31 GiB GPUs (RTX 5090):

1. _patch_cpu_offload_apply now intercepts .to(cuda:X) in addition to .cuda()

   • Callers using model.to(device) (common in multi-model pipelines) bypassed the offload hook entirely, putting all weights on GPU and OOMing on 31 GiB cards.
   • Fix: probe the .to() lambda with a small CPU tensor to detect CUDA target.

2. nd_tiling max_tiles capped at 2 unconditionally

   • max_tiles=3 overflows the CUDA z-grid limit (65535) on conv-heavy dynamic-shape graphs (e.g. 1080p VAE decode). Grid3D has no z-overflow handling.
   • Previously only guarded on PT>=2.12; now capped on all versions.

3. max_tiles exposed as MAGI_COMPILE_PASS_CONFIG__ND_TILING_MAX_TILES

   • Allows experimentation without code edits. Default remains 2.

4. Prefetch lookahead configurable via OffloadConfig.max_prefetch_lookahead

   • Hard-coded 2 layers of lookahead costs GPU memory a small card may not have. Setting to 0 disables prefetch entirely.

[jiahy0825]

LGTM