Fix device mismatch in measure_peak_memory: move model to device, guard CUDA peak-stat calls, capture loss

Replace mispredicting memory estimate with a measured preflight: real fwd+bwd on the longest example, auto-select max_seq_len

Fix CUDA OOM: max_seq_len 2048->1024, enable expandable_segments, add memory preflight calibrated to the observed 13.5GB failure

Support transformers 5.x: TrainingArguments adapter converts warmup_ratio to warmup_steps; drop unsupported kwargs with a warning; test covers both

Fix: resolved_precision now accepts the (major, minor) tuple from torch.cuda.get_device_capability; add tests module for GPU-only paths

Initial: pipeline, training, eval modules