Pass chunk size to moe op

[yiliu30]

Requires HabanaAI/vllm-hpu-extension#337

[Copilot]

Pull Request Overview

This PR modifies the FP8 quantization MoE (Mixture of Experts) forward pass to support passing chunk size information to the underlying MoE operation. The change enables dynamic chunk size configuration by extracting tokens_num from hidden_states and delegating to the original module for additional kwargs.

• Adds a helper method to extract extra kwargs from the original module
• Modifies forward_quant to pass chunk size information via extra_kwargs

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

This assumes hidden_states has exactly 2 dimensions, but tensors could have more dimensions (e.g., batch_size, sequence_length, hidden_dim). Consider using hidden_states.shape[0] or hidden_states.shape[-2] depending on the expected tensor layout.

Suggested change

	tokens_num, hidden_dim = hidden_states.shape
	tokens_num = hidden_states.shape[-2]
	hidden_dim = hidden_states.shape[-1]