add 9-config hparam sweep + new_layer_lr_mul param-groups support

fix scheduler bug: don't prepare scheduler with accelerate (was over-stepping cosine by num_processes); add grow40_winning_v2 config

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