[OpenVINO] Add Qwen3.5 (Gated Delta Networks) export and inference support

[taowen-paraflow]

Summary

• Add complete OpenVINO export and stateful inference support for Qwen3.5 (Gated Delta Networks hybrid architecture)
• Qwen3.5 combines 18 linear attention layers (GatedDeltaNet with conv_states + recurrent_states) and 6 full attention layers (standard KV cache)
• Include compatibility fixes for transformers 5.x and OpenVINO dev version parsing

Qwen3.5 core support (6 files)

File	Change
model_configs.py	Qwen3_5OpenVINOConfig + Qwen3_5DummyPastKeyValuesGenerator for hybrid cache (conv + recurrent + KV)
model_patcher.py	Qwen3_5Patcher wrapping Qwen3_5DynamicCache, patching GDN layers to torch fallback paths
utils.py	Register qwen3_5 / qwen3_5_text in SSM_MODELS and SKIP_CHECK_TRACE_MODELS
stateful.py	patch_stateful_hybrid_ssm supports recurrent prefix for GDN states
modeling_decoder.py	OVCacheWithMambaStates extended with recurrent_states; token-by-token prefill with position_ids
configuration.py	Default int4 quantization presets for Qwen3.5-3B and Qwen3.5-8B

Compatibility fixes (7 files)

File	Change
import_utils.py	OpenVINO dev version string parsing ("2026.0.0-17740-abc" → "2026.0.0")
modeling_base.py	AttributeError handling for composite configs; is_offline_mode import fallback
modeling_open_clip.py	is_offline_mode import fallback
modeling_seq2seq.py	AutoModelForVision2Seq import fallback (removed in transformers 5.x)
utils.py (intel)	ParameterFormat / compute_serialized_parameters_size inline fallback
modeling_utils.py	HfFolder shim for huggingface_hub 0.25+
setup.py	Remove transformers<4.58 upper bound (Qwen3.5 requires ≥5.3)

Key design decisions

• Decode-only export: Model is traced with seq_len=1; prefill is handled token-by-token at runtime with correct position_ids for RoPE
• No CUDA dependencies: GDN layers are patched to use torch fallback paths (torch_causal_conv1d_update, torch_recurrent_gated_delta_rule) during export, avoiding flash-linear-attention / causal-conv1d CUDA kernels
• OV native stateful conversion: Uses apply_make_stateful_transformation directly (no custom reimplementation)
• 48 stateful variables: 18 conv + 18 recurrent + 6 key + 6 value for the 0.8B model
• bfloat16 handling: Added bf16→OVType.bf16 path in _get_input_info since NumPy lacks bf16 support

Tested on

• Qwen3.5-0.8B on Intel Core Ultra 7 258V (Lunar Lake), CPU inference ~6-10 tok/s
• Output quality verified against PyTorch reference ("The capital of France is Paris", arithmetic patterns, etc.)

Test plan

• Export Qwen3.5-0.8B to OpenVINO IR with optimum-cli export openvino
• Run stateful inference and verify output quality
• Verify no regression on existing SSM models (Zamba2, Mamba, etc.)
• Add tiny-random-qwen3.5 model to HF Hub for CI testing

🤖 Generated with Claude Code

[rkazants]

Hi @taowen-paraflow, thanks a lot for your contribution but we already on this task in this PR: #1634

Feel free to take good-first-issue for contribution.