fix(qwen35): trim legacy device pools between requests

[boxwrench]

Summary

• add an explicit trim hook for legacy CUDA/HIP device pools
• release per-evaluation Q8 memo allocations before trimming cached blocks
• trim target and distinct draft backends when Qwen3.5 request scratch is released
• extend the existing CUDA pool-shutdown regression test through explicit trimming

Root cause

On non-VMM ROCm devices, long prefills returned operator temporaries to the legacy ggml pool, but the pool retained those blocks indefinitely. A later request with a different allocation shape could require another large contiguous allocation and fatally OOM even though the previous request had released its scratch tensors.

Validation

• HIP gfx1201 release build completed for dflash_server and test_server_unit
• two independent back-to-back requests completed on an AMD Radeon AI PRO R9700: 40,013 and 39,234 prompt tokens
• request cleanup returned 2,405.3 MiB and 2,243.6 MiB from device pools respectively
• unit suite: 421/423 passed in the restricted workspace; the two socket-permission failures passed when rerun with normal socket permissions
• git diff --check clean

[cubic-dev-ai]

All reported issues were addressed across 5 files

_{You’re at about 90% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[davide221]

@boxwrench thanks for the contribution! We have fixed cubic issues.