vulkan: decline GATED_DELTA_NET raw gates instead of computing them wrong

[bri-prism]

Fixes #237.

Problem

ggml_gated_delta_net_set_raw_gates() delivers beta and g pre-activation, so the op has to apply

beta = 1/(1 + exp(-beta))                    // sigmoid
g    = a[h] * softplus(g + dt_bias[h])       // dt_bias = src[7], a = src[8]

as the CPU implementation does (ggml/src/ggml-cpu/ops.cpp, the if (raw_gates) branch).

gated_delta_net.comp has neither binding and does none of that math — it applies exp(g) unconditionally. But supports_op only checks src[6], the head size and the src types; it never looks at op_params[1]. So the Vulkan backend claims these ops and returns results uncorrelated with the reference, silently — no abort, no warning, no fallback.

Measured on an Intel Arc B390 before this change, test-backend-ops test -o GATED_DELTA_NET -b Vulkan0:

case (head_count=4, head_size=128)	ERR (tolerance 1e-7)
n_seq_tokens=1, n_seqs=1, v_repeat=1	0.897 - 0.996
n_seq_tokens=1, n_seqs=2, v_repeat=2	0.958 - 0.992
n_seq_tokens=64, n_seqs=1, v_repeat=1	1.000

15/15 failures over five consecutive runs. All 45 raw_gates=0 cases pass, and no raw_gates=1 case has ever passed — so the split is exactly on this flag. The SYCL backend on the same machine declines all four raw_gates=1 cases as not supported, which is the behaviour this change adopts.

Change

Nine lines in supports_op: decline when ggml_get_op_params_i32(op, 1) != 0, with a comment recording why.

This makes the case fall back to the CPU, which computes it correctly. It does not implement raw gates on Vulkan — that needs two new shader bindings plus the sigmoid/softplus math and care around the kda path, which is a much larger change. The trade here is speed for correctness on those shapes, which seems clearly right when the current behaviour is silent corruption.

Result

test-backend-ops test -o GATED_DELTA_NET -b Vulkan0, three consecutive runs: 36/36 passed, Backend Vulkan0: OK each time. The four raw_gates=1 cases now report not supported [Vulkan0].

Full unfiltered test-backend-ops test -b Vulkan0 on the same device:

	OK	failing cases	not supported	verdict
prism 422590f5d	17187	3	4667	FAIL
this branch	17188	0	4670	OK

The three failures become not supported; the Vulkan backend passes the suite on this device for the first time.

Note for #187

PR #187 implements the rows-indexed state path, which makes supports_op accept a fourth raw_gates=1 case (K=2, rows_mode=1) that stock prism declines for an unrelated reason. That case then executes and hits this same bug, which looks like a regression in #187 but is not one — it is this bug surfacing on one more shape. With this change applied that case is declined too, and #187's GDN results are clean.