[jvmenen]

Fixes #4725.

gated_delta_update cast g and beta to the dtype of q before the kernel. With bf16 activations a decay just below 1 rounds to 1.0 (0.999, 0.9995 and 0.9999 all do), so the state stops decaying at that token. The kernels already convert both to float32 right after loading; only the round trip before the kernel lost the precision.

g and beta are now cast to float32, and the forward and VJP kernels read them as const device float*. The kernel template type stays that of q, k and v, so there are no new instantiations. Float32 inputs give the same result as before; dg and dbeta are cast back to the caller's dtype by the astype in front of the primitive.

Error compared to a float64 result, for a Qwen3.5-sized layer on an M5 Pro:

inputs	output error	final state error	time
bf16 q/k/v, main	4.8e-2	5.0e-2	4.91 ms
all float32 (current workaround)	1.8e-3	1.3e-3	6.45 ms
bf16 q/k/v, this PR	2.4e-3	1.3e-3	5.20 ms

(4096 tokens, 16 key and 48 value heads, head dim 128, decay values as in a trained model.)

test_gated_delta_half_precision_decay_near_one runs bf16 and fp16 q/k/v with decays between 0.99 and 0.9999 against the same call on float32 inputs, on CPU and on the three GPU routes. It fails on v0.32.4 (state error 2.3e-2) and passes with this change.