[zcbenz]

Add a simple CUDA kernel for gated_delta_update that updates states sequentially, I will add a chunked kernel in follow up PR.

This kernel is modified from the submission of @kamahori from MLSys 2026 Competition, and the modification was done by myself instructing AI.

Benchmarks by running benchmarks/python/gated_delta_bench.py:

B	T	Hk	Hv	Dk	Dv	fallback (ms)	seq (ms)	seq (speedup)
1	8	16	32	128	128	2.239	0.049	45.52x
1	32	16	32	128	128	9.161	0.104	88.14x
1	64	16	32	128	128	18.124	0.125	144.53x
1	128	16	32	128	128	38.162	0.206	185.11x
1	256	16	32	128	128	79.476	0.495	160.55x
1	512	16	32	128	128	164.421	0.827	198.74x
8	8	16	32	128	128	7.096	0.210	33.73x
8	32	16	32	128	128	25.544	0.482	52.98x
8	64	16	32	128	128	51.001	0.721	70.72x
8	128	16	32	128	128	101.674	1.419	71.65x
8	256	16	32	128	128	203.280	2.360	86.13x
8	512	16	32	128	128	406.669	4.646	87.53x

[ArupKantiDas]

I ran this on an M5 Pro (applegpu_g17s, which has NAX), since none of the CI runners have NAX.

• test_fast_gated_delta.py passes: whole file, -n 4, each test alone, and with MLX_ENABLE_TF32=1.
• I also checked that the NAX kernels actually run under the new scoped configs. With MLX_GATED_DELTA_CHUNK=16 the forward is about 2.5e-2 off the masked fallback, the same gap the NAX kernel shows on main. The chunk=16 VJP differs from the sequential one too (max dq difference 2.3e-4, against 1e-7).
• Full suite: 974 passed, 1 failed. The failure is test_conv test_input_dilation_grad, which fails on main here as well and passes when run alone.
• Heads-up for Respect MLX_ENABLE_TF32 in gated_delta_update #4792: it conflicts with this PR in gated_delta_update.cpp, because both rewrite the chunk-size helpers. I merged them locally (this PR's MLX_ names plus Respect MLX_ENABLE_TF32 in gated_delta_update #4792's dtype/TF32 check) and got the same test results, and the mx.fast.gated_delta_update VJP: wrong gradient w.r.t. g (and ~1e-3 error on all inputs) on the NAX path for Hk=16 #4614 repro stays fixed with TF32 off (dg rel. L2 2.1e-07).
• Small thing: GATED_DELTA_CKPT (gated_delta_update.cpp#L30) keeps its old name, while the other four keys got the MLX_ prefix. Is that on purpose?

I used Claude Code to build and run the tests above.