Add manual_attention_part_hp (fp32-logits variant)

Add manual_attention_part_v2 (bf16-saved-p, fused softmax gradient)

Add package source and torch-universal build variant