opencl: make the MoE expert scatter deterministic

[wanghqc]

Overview

This PR is to fix a bug in MoE scatter.

• MoE prompt processing on the OpenCL backend is not reproducible run-to-run: the same binary on the same prompt returns one of several outputs.
• Measured on granite-3.0-3b-a800m-Q4_K_M (Adreno X2-90, greedy, llama-simple -n 64, 3 prompts × 5 runs each), every prompt produced 2–3 distinct outputs across 5 runs.

Root cause

• kernel_moe_scatter assigns each token its slot within its expert via atomic_inc(&slot_counter[expert]), so the token→slot packing depends on which work-item wins the atomic and varies every run.
• The ragged MoE prefill GEMM is sensitive to that packing order; the non-ragged path happens to be immune.

Fix

• This PR replaces the atomic slot grab with a deterministic rank: a token's slot is its position in flat (token, k) order among the tokens routed to the same expert — a fixed function of the routing input.
• One workgroup per expert walks the routing list in blocks of 64 with a workgroup scan and a running carry.
• Default on; GGML_OPENCL_MOE_STABLE_SCATTER=0 restores the previous behavior.
• 2 files, backend-only (ggml-opencl.cpp + kernels/moe_sort_by_expert.cl).

Additional information

Determinism verification

• With the fix: 5/5 identical runs on every prompt, and the output matches an O(N·topK) reference implementation exactly.
• Six MoE models on the X2-90, 3 runs each, all single-output and coherent: gpt-oss-20b-MXFP4, Qwen3-30B-A3B-Q4_K_M, Qwen3-unsloth-30B-A3B-Q4_0, Qwen3.6-35B-A3B-MXFP4_MOE, and granite-3.0-3b in Q4_K_M and MXFP4_MOE.
• gpt-oss also 3/3 on the Adreno 840.

Cost

Prefill only — the scatter runs on the GEMM path, decode never calls it: pp512 −0.63% on granite-3.0-3b (40 experts), −0.15% on Qwen3-30B-A3B (128 experts).

Scope

• This fixes run-to-run reproducibility.
• The ragged and non-ragged paths still produce different roundings of equal quality (MUL_MAT_ID max err 2.68e-05 vs 2.79e-05 against the CPU reference);
• A host comment claiming the ragged path is byte-identical is corrected as part of this change. That difference is a rounding choice, not a wrong answer.

Testing

test-backend-ops -o MUL_MAT_ID: Adreno X2-90, 383 OK / 0 FAIL.
The Adreno 840 declines all MUL_MAT_ID cases at these shapes in test-backend-ops; its evidence is the real-model determinism runs above.

Requirements

• I have read and agree with the contributing guidelines Yes
• AI usage disclosure: Yes, testing.