Add Qwen3.5 support (hybrid gated-DeltaNet + attention decoder): dense 2B/27B, then sparse 35B-A3B

Summary

I have Qwen3.5 running on this plugin at release-0.24.0.1.1.0 — the dense
checkpoints (2B, 27B) and, separately, the sparse 35B-A3B — and would like to
contribute them upstream. Opening this first per CONTRIBUTING ("open an issue to
discuss any significant work"), because the dense half touches shared code and
I would rather find out now if the approach is wrong.

What Qwen3.5 needs that the plugin does not have yet

Qwen3.5's decoder is hybrid: most layers are gated DeltaNet, a linear
recurrence carrying fixed-size state per sequence, and the rest are ordinary
full attention (18 + 6 on 2B, 48 + 16 on 27B). vLLM puts the two kinds of layer
in separate KV cache groups, which this plugin has not had to support before.
Four things follow:

1. A model needs a way to declare state-carrying layers. KVSpec/LayerSpec
   describe paged attention layers only; num_kv_heads/head_size are
   meaningless for a recurrent layer, where what matters is the concrete state
   tensor shapes.

2. Page sizes have to be reconciled across the two groups.
   unify_kv_cache_spec_page_size requires every group's page size to divide the
   largest. A DeltaNet state page is 271360 bytes per rank at TP=4, which factors
   as 1024 * 5 * 53, so no sane attention block size divides it and startup
   fails with "Cannot unify by adjusting block_size". vLLM solves this in
   Platform._align_hybrid_block_size — grow the attention block size, then pad
   the state page to match — but it gates that on _find_non_ssm_backend
   finding one of its own attention backends, and NeuronPlatform registers
   none, so it never runs here.

3. Neuron device tensors reject vLLM's state layout. gpu_model_runner
   strides each state tensor out of the layer's raw buffer with the block dim
   striding by a whole page. On Neuron that raises Detected non-contiguous slicing for requested Device Tensor for any view whose stride is not the
   natural one, so the layout has to be chosen differently — and the choice
   interacts with whether the raw buffer is shared across groups.

4. --modular-flow-mac-threshold=10 breaks codegen on a pure-torch graph.
   The runner passes it unconditionally. On this model's decode graph neuronx-cc
   fails with NCC_IBTN006 (a pftranspose whose copy fails backend
   verification). Confirmed narrowly by recompiling the cached graph.hlo by
   hand: fails with the flag, compiles without it. The threshold exists because
   NKI kernels do not report MAC counts, and the dense path calls no kernels.

What I propose

A PR against release-0.24.0.1.1.0 (not main — the code is written against
0.24.0.1.1.0's neuron_model_runner/platform surface, and recent PRs target
that branch) adding:

• vllm_neuron/model/qwen3_5/ — the model, and Qwen3_5ForConditionalGeneration
  in the registry.
• RecurrentLayerSpec + KVSpec.recurrent_layers in model/kv_cache.py.
• MambaSpec construction and both cache layouts in the model runner.
• _align_hybrid_page_sizes in platform.py, which calls vLLM's own helper with
  a stub backend supplying the one thing it reads, rather than reimplementing the
  arithmetic — the same helper sizes the state pages the planner allocates, and a
  second copy of that arithmetic is a silent aliasing bug waiting to happen.
• NeuronConfig.hlo2tensorizer_options, so a model can opt out of the threshold.
• examples/vllm_neuron/models/qwen3_5/ with a run script, a latency benchmark
  and an HF-comparison accuracy check.

Measured on a trn2.3xlarge at TP=4, bf16, greedy, 1024 in / 128 out: 2B does
TTFT 145 ms / TPOT 3.88 ms at batch 1 and 465 output tok/s aggregate at batch 8.
27B runs with gpu_memory_utilization=0.65 and -O3; the reason for each is
documented in the example README, along with the known limitations (prefix
caching must stay off; one batch bucket per process; decode gathers the full
block-table width rather than only the occupied pages).

The sparse 35B-A3B support is a second, independent PR — a Qwen3_5SparseMoeBlock
plus a second registry entry — so that the dense change can be reviewed on its
own.

Attribution

The Qwen3.5 implementation derives from
qingzwang/vllm-neuron branch
model/Qwen3.5-2B (Apache-2.0), which is where the 2B port originates;
nki_deltanet.py, nki_deltanet_fused.py, vl.py and flags.py are that
fork's files unchanged, and the bulk of model.py, deltanet.py and config.py
is its work. My contribution is the forward-port to 0.24.0.1.1.0, the 27B
checkpoint, and the sparse variant. Credited in the commit trailer as well.

Questions

1. Is release-0.24.0.1.1.0 the right base, or would you prefer main?
2. Is extending KVSpec with a separate recurrent_layers list the shape you
   want for hybrid models, or would you rather LayerSpec grew a kind
   discriminator?
3. There is no test directory on this branch. Where should model tests go?

Happy to split, rebase or restructure — please say which.

Start on release-0.24.0.1.1.0 by reviewing model/kv_cache.py, platform.py, neuron_model_runner, and NeuronConfig.hlo2tensorizer_options. Compare the proposed hybrid cache and page-size handling with vLLM's _align_hybrid_block_size, then inspect the listed qwen3_5 entry points and examples. Done means dense 2B and 27B support is registered, runnable, and covered by the proposed accuracy and benchmark checks.