Build a block and verify parameter counts against real models

Two independent pieces: block.py implements the block, and params.py counts parameters from a config without building anything. They meet in one check: the block's real parameter count must equal the formula. The decision that matters most is checking each formula against a real implementation, not only against a quoted total: GPT-2 against GPT2LMHeadModel, and the Llama formula against a tiny LlamaForCausalLM with grouped-query attention, where a wrong K/V width would show immediately.