grow40_winning: switch student to bf16 to fit in B200 memory + 40-layer Adam state

add retry loop around load_dataset for transient HF Hub 5xx

add micro_batch_size config key + per-micro inner loop in train step (fixes OOM for fp32+seq2048)

fix OOM: chunked KL with checkpointing + PYTORCH_CUDA_ALLOC_CONF expandable_segments; add kl_chunk_size config key

add grow_layers, sweep configs (replicate_zero4, grow40_winning, grow40_simple), sweep runner

initial scaffold: distill.py + base/zero_14_17 configs + accelerate yaml

initial commit