[rohith500]

Fixes #138

Two fixes for _FP16Casting:

1. Guards torch.topk on ANE: Running aten.topk in FP16 on Apple Neural Engine packs two 16-bit indices into each 32-bit output word (Apple Radar FB24904516 / FP16 model silently returns wrong outputs with default specialization on iPhone 17 Pro; CPU-only, GPU-preferred and FP32 are correct coreai-torch#115), corrupting downstream indexing in detection models like RF-DETR and YOLO. Added guard_ane_topk: bool = True (with opt-out) to keep topk in FP32 with boundary casts.

2. Downcasts multi-output containers at operator.getitem: Multi-output operations returning tuples (topk, split, chunk) that ran in FP32 skipped downcast insertion because container tuples lack .dtype. Extracted FP32 tensors leaked into downstream layers, forcing them into FP32 and triggering runtime weight upcasts. We insert downcasts on extracted float elements in operator.getitem (preserving FP32 if the activation overflowed during calibration or was explicitly ignored).

Tested with pytest tests/casting/ -v (142 passed) and make check (all 25 checks passed).