Add DSpark drafter for Gemma 4 (#2137)

Backbone variant of the model-agnostic dspark drafter: reuses the shared
block sampling, confidence head and dflash round loop, and swaps in Gemma 4
draft layers (global_head_dim RoPE, attention_k_eq_v, sandwich norms,
layer_scalar).

Adds a _project_kv hook on DFlashAttention so a checkpoint that shares one
projection for keys and values does not need a forked attention block.