[Model][Quantization] Fix / Add GGUF support for Qwen2 MoE models

[a4lg]

Purpose

This is a follow-up after #30116 (which unblocks loading Qwen2/3 MoE + GGUF models).

This commit fixes two issues preventing to load Qwen2 MoE + GGUF models:

1. Initialization of Qwen2MoeModel.embed_tokens is fixed to use correct prefix and respect quantization settings.
2. Added GGUF-specific compatibility layer for Qwen2MoeSparseMoeBlock.shared_expert_gate (GGUF: 1D tensor (n), HF/vLLM: 2D tensor (1, n)).

The latter was implemented in HF Transformers as a part of Qwen2MoeTensorProcessor but since vLLM's weight loader directly uses gguf library without such compatibility layer, we need to implement the equivalent.

Test Plan

You may download a quantized Qwen1.5-MoE-A2.7B-Chat model such as tensorblock/Qwen1.5-MoE-A2.7B-Chat-GGUF and run like:

vllm serve \
	Qwen1.5-MoE-A2.7B-Chat-Q4_K_S.gguf \
	--tokenizer Qwen/Qwen1.5-MoE-A2.7B-Chat \
	--max-model-len 4096 \
	--gpu-memory-utilization 0.5

WARNING: Qwen2-57B-A14B-Instruct with the same architecture won't work (alone) because a bug in HF Transformers prevents loading non-default parameters.
See #30116 for details (fixed in upcoming V5 by huggingface/transformers#42650 but not sure whether V4 (which vLLM currently depends on) will be fixed the same).

Test Result

Before this PR, either one of the errors will occur (each corresponds to a fix above):

1. KeyError: 'embed_tokens.qweight_type' (you should normally see this) or
2. AssertionError: Attempted to load weight (torch.Size([2048])) into parameter (torch.Size([1, 2048]))
   (2048 is the value on Qwen1.5-MoE-A2.7B{,-Chat}).

After this PR is applied, these errors will go away and you should be able to use Qwen2 MoE + GGUF models.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request introduces fixes to enable GGUF support for Qwen2 MoE models. The changes include correctly initializing embed_tokens to respect quantization settings and adding a compatibility layer to handle a shape mismatch for the shared_expert_gate weight in GGUF models.

The changes are correct and address the described issues. I have one suggestion to make the weight loading logic for shared_expert_gate more robust by adding more explicit checks against the model's parameter shape, which will prevent potential issues with other formats in the future.