fix: remove hardcoded cuda:0 in OTA loss for multi-GPU support

[Mr-Neutr0n]

Bug

The OTA loss computation hardcodes cuda:0 for tensor allocation, causing failures during multi-GPU (DDP) training when the model runs on other GPUs.

In utils/loss.py, several locations use .cuda() or device='cuda:0' when creating tensors, which forces them onto GPU 0 regardless of where the model and input data reside. This causes device mismatch errors during distributed training.

Fix

Replaced hardcoded device references with dynamic device inference from input tensors:

• .cuda() → .to(logits.device) in RankSort, aLRPLoss, and APLoss forward methods
• device='cuda:0' → device=targets.device in build_targets and build_targets2 methods

This is consistent with how device handling is already done elsewhere in the same file (e.g., torch.ones(7, device=targets.device)).