Fix Qwen3OmniMoe Talker loading and config initialization

[Krish2002]

What does this PR do?

This PR fixes two issues that prevented the model
Qwen/Qwen3-Omni-30B-A3B-Instruct from loading correctly with
AutoModelForMultimodalLM.

---

Fix 1: AttributeError: Qwen3OmniMoeTalkerForConditionalGeneration has no attribute 'lm_head'

Issue

Qwen3OmniMoeTalkerForConditionalGeneration deletes lm_head in its __init__
method, but it inherits _tied_weights_keys from its parent class
(Qwen3MoeForCausalLM), which references lm_head.weight.

During model loading, mark_tied_weights_as_initialized() attempts to access
lm_head.weight, resulting in an AttributeError.

Fix

Explicitly set _tied_weights_keys = {} for
Qwen3OmniMoeTalkerForConditionalGeneration, since this model does not use tied
weights.

This change is implemented in:

• src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py

---

Fix 2: AttributeError: ... object has no attribute 'initializer_range'

Issue

Several composite config classes were missing the initializer_range attribute:

• Qwen3OmniMoeTalkerConfig
• Qwen3OmniMoeCode2WavConfig
• Qwen3OmniMoeConfig

When _initialize_missing_keys() runs during model loading, it may call
_init_weights() for modules that are not loaded from the checkpoint.
_init_weights() expects self.config.initializer_range, which caused an
AttributeError.

Fix

Add an initializer_range parameter (default: 0.02) and store it as an
attribute in the affected config classes.

This change is implemented in:

• src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py

---

Breaking changes

None.

---

Tests

Added a unit-level regression test in:

tests/models/qwen3_omni_moe/test_configuration_and_loading.py

The test verifies that:

• Qwen3OmniMoeTalkerForConditionalGeneration has empty tied weight keys.
• Qwen3OmniMoeTalkerConfig, Qwen3OmniMoeCode2WavConfig, and
  Qwen3OmniMoeConfig all define the initializer_range attribute.

Additionally, verified locally that the model loads successfully using the
following reproduction script:

from transformers import AutoModelForMultimodalLM
import torch

model = AutoModelForMultimodalLM.from_pretrained(
    "Qwen/Qwen3-Omni-30B-A3B-Instruct",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: qwen3_omni_moe

[Rocketknight1]

cc @zucchini-nlp

[zucchini-nlp]

resolved in #43084, closing as duplicate