Add dense Qwen3.5/3.8-27B checkpoint-conversion support

[lokic233]

What

Adds HF→MaxText checkpoint-conversion support for the dense Qwen3.5/3.8-27B model.

Why

The qwen3_5 decoder block currently assumes the MoE variant throughout:

• param_mapping unconditionally registers routed_experts-* and shared_expert-* hooks
• hf_shape derives MLP shapes from the expert config
• there is no qwen3.5-27b model config at all

The 27B checkpoint is dense — num_experts is absent from its config.json
entirely — so conversion currently emits hooks for parameters that do not exist and
misses the three that do.

Worth flagging for reviewers: Qwen/Qwen3.8-27B declares model_type: qwen3_5 and
architectures: ["Qwen3_5ForConditionalGeneration"], so the existing qwen3_5 block is
the correct home for it despite the version number in the name.

Changes

• configs/models/qwen3.5-27b.yml (new) — dense 27B: 5120 hidden, 64 layers,
  24 q / 4 kv heads, head_dim 256, ffn 17408, vocab 248320,
  inhomogeneous_layer_cycle_interval: 4 (one full-attention layer per four),
  GDN 16 key / 48 value heads at 128 dim, conv kernel 4,
  rope_max_timescale: 10000000, partial_rotary_factor: 0.25.
• param_mapping.py — gate the expert hooks on num_experts > 1; add the dense
  mlp-wi_0 / wi_1 / wo branch. The MoE shared_expert is the same MlpBlock class the
  dense branch instantiates, so the same plain transpose applies.
• hf_shape.py, hf_model_configs.py — dense MLP shapes, and the 27B HF config entry.
• models/qwen3_5.py, configs/types.py, utils/globals.py — plumb the dense path
  and register the model name.

Every dimension is transcribed from Qwen/Qwen3.8-27B's own config.json, verified
byte-identical by md5 (0ecc077f5c9eb49889aa0022f4ae9fa7) against the stored artifact.

Testing

Hook-table diff, driving QWEN3_5_MAXTEXT_TO_HF_PARAM_HOOK_FN with the real
Qwen3.8-27B config.json:

	hooks	expert hooks	dense MLP hooks
before	769	512 (for parameters that do not exist)	0 of 3
after	449	0	192 (64 layers × wi_0/wi_1/wo)

The MoE arm is the control. Forcing num_experts: 8 into the same config yields an
identical 769-hook table before and after, with the same eight -mlp- hook kinds — the
gate does not disturb the path it gates. Attention hooks are unchanged in both arms.

Numerical check against the transformers reference on CPU at reduced dimensions
(hidden 256, 8 layers, vocab 512, every parameter randomized — default init leaves
RMSNorm scales at 0, which would hide a wrong norm mapping):

	max|diff|	cosine	argmax agreement
scan_layers=False	1.66893e-06	1.00000000	100%
scan_layers=True	1.57952e-06	1.00000000	100%

Sensitivity is demonstrated rather than assumed: swapping gate_proj↔up_proj — a
same-shape error no assertion in the converter can catch, since silu(g)*u != silu(u)*g
— scores cosine 0.72936636 / 18.8% argmax agreement.

Depends on #5148

At --simulated_cpu_devices_count=1 this feature cannot run without the Orbax v1
leaf-handler fix (#5106); conversion dies in the save path before producing anything.
It works at the default of 16 either way. Please land that one first, or this ships a
feature that is dead on the single-device path.

Rebased onto 4cfec5b20c26fa2775e760034b66b7a4645ba877.

[gemini-code-assist]

Code Review

This pull request adds support for the dense qwen3.5-27b model (Qwen3.8-27B) to the Qwen3.5 model family. It updates the configuration, shape utilities, parameter mapping, and model architecture to support both dense and MoE variants dynamically. Feedback suggests handling cases where 'num_experts' is explicitly set to None in the configuration dictionary to prevent a potential TypeError.

[gemini-code-assist]

To prevent a potential TypeError if "num_experts" is explicitly set to None in the configuration dictionary, handle None values defensively. This aligns with the defensive style used in GEMMA4_HF_WEIGHTS_TO_SHAPE in this same file.

Suggested change

	# Dense members of the Qwen3.5 family (e.g. qwen3.5-27b) omit every MoE key.
	num_experts = config.get("num_experts", 1)
	is_moe = num_experts > 1
	# Dense members of the Qwen3.5 family (e.g. qwen3.5-27b) omit every MoE key.
	num_experts = config.get("num_experts")
	num_experts = num_experts if num_experts is not None else 1
	is_moe = num_experts > 1