[Vulkan] Add multi-queue decode correctness and queue-ownership stress suite

[goniz]

Problem

As the Vulkan backend becomes more asynchronous and decode-oriented, generation correctness needs dedicated stress coverage for queue ownership, timeline sequencing, and cross-queue handoff. The current multi-queue work is not sufficiently protected by targeted decode-focused correctness tests.

This is not a duplicate of #1. Issue #1 added the multi-queue submission model. This issue is about proving it stays correct under real decode-like workloads and future performance changes.

Why This Matters

The reference review reinforced that MLX should keep a stronger correctness model than the reference runtimes, especially for cross-queue synchronization and queue-family ownership. If we reduce synchronization to improve performance, we need tests that catch subtle generation regressions early.

Tasks

• Add targeted tests for same-family dual-queue setups and separate-family compute/transfer setups
• Stress upload -> compute -> KV append -> readback ordering in decode-like workloads
• Add assertions or trace validation around timeline values and queue-affinity transitions where practical
• Add short-timeout reproduction harnesses for generation correctness under multi-queue execution

Acceptance Criteria

• Dedicated multi-queue decode correctness coverage exists in CI or reproducible local tests
• Same-family and separate-family queue topologies are both exercised where available
• Future sync/perf changes can be validated without relying only on large manual benchmark runs

References

• mlx-vulkan-reference-conclusions.md
• references/ggml-vulkan-findings.md
• references/zinc-findings.md
• references/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp
• references/zinc/src/vulkan/command.zig
• references/zinc/src/compute/forward.zig

Related

• [Vulkan] Implement multi-queue backend with timeline semaphore submission model #1