[Medamineelkhattabi]

Follow-up to #1115.

In the GatedDeltaNet code in ch04/08_deltanet/README.md, W_beta maps to d_out, so the write strength β is computed per channel. Qwen3-Next and the Gated DeltaNet paper use one β per head (in_proj_b outputs num_v_heads, then beta = b.sigmoid() in modeling_qwen3_5.py).

This changes W_beta to nn.Linear(d_in, num_heads) and keeps β as (b, num_tokens, num_heads, 1), so delta = (v_t - kv_mem) * b_t broadcasts over the head dimension. The recurrence loop and the snippet further down the README are unchanged.

I checked it by running the README class and comparing its recurrence with HF's torch_recurrent_gated_delta_rule on random inputs (same q, k, v, decay and β): max abs difference 3.7e-9. With the previous per-channel β the two can't be compared directly, since the shapes differ.