[mem_cache] Run mamba models on UnifiedRadixCache when the radix cache is disabled

[hnyls2002]

• With --disable-radix-cache, mamba models (GDN, Mamba2, KDA, lightning, incl. MiniCPM SALA and the PD decode side) now use UnifiedRadixCache's disabled mode instead of ChunkCache; the mamba component owns mamba-state release in both modes.
• MambaComponent requires page_size == 1 only when the tree caches states; a disabled tree caches none. This also fixes --disable-radix-cache --enable-streaming-session on mamba models, which crashed at startup since [Session] Run streaming sessions only on UnifiedRadixCache; reject unverified tree caches #42295.
• Gate the mamba COW headroom (allocation.py) and exact_chunk_fill on the tree actually caching mamba states.
• MLX: drop the pool-owned auxiliary-state release path; MlxAuxiliaryStateComponent releases in both modes.
• Reject mamba models on tree caches that do not manage mamba state (e.g. --enable-flexkv, third-party backends) as not verified, and drop the release_kv_cache fallback that freed mamba state for them.
• Behavior change: beam search now rejects mamba models with --disable-radix-cache as well (it already rejected them with the radix cache on). Previously ChunkCache reported supports_mamba() == False and slipped past the check, but beam forking copies only the KV page mapping (beam_search/fork.py) and never the mamba state, so those beams ran on the wrong recurrent state.

Verification (Qwen3.5-4B, --disable-radix-cache, strict busy mem check + KV page invariants + tree sanity check on): GSM8K 400q default / retract / no-overlap 0.8775 / 0.8875 / 0.8825 (main with ChunkCache: 0.875-0.8925), concurrent abort storms, streaming-session kit (default, retract) and abort-leak repro all pass with no leak. Output throughput with checks off matches main within noise (avg 2692 vs 2627 tok/s over 3 alternating pairs).

---

CI States

Latest PR Test (Base): 🚫 Run #37115199321
Latest PR Test (Extra): ⚠️ Not enabled -- add run-ci-extra label to opt in.
Latest PR Test (AMD ROCm 10): ❌ Run #37115199488

[hnyls2002]

/rerun-test test_disaggregation_hybrid_attention.py test_qwen35_gdn_multi_item_scoring.py test_nvidia_nemotron_3_super_bf16_mtp.py test_streaming_session.py test_mamba2_extra_buffer_kl.py

[github-actions]

Results for /rerun-test test_disaggregation_hybrid_attention.py test_qwen35_gdn_multi_item_scoring.py test_nvidia_nemotron_3_super_bf16_mtp.py test_streaming_session.py test_mamba2_extra_buffer_kl.py:

🚀 8-gpu-h200 (2 tests): ✅ View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_hybrid_attention.py
cd test/ && python3 registered/e2e/models/test_nvidia_nemotron_3_super_bf16_mtp.py

🚀 1-gpu-h100 (3 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/gdn/test_qwen35_gdn_multi_item_scoring.py
cd test/ && python3 registered/sessions/test_streaming_session.py
cd test/ && python3 registered/radix_cache/test_mamba2_extra_buffer_kl.py