[refactor]support gatingtopk operator generalization

[1092626063]

What this PR does / why we need it?

This pr is cherry-pick from : #2958 and #4340

Past：
npu_moe_gating_top_k can only support 'group_count=256' pattern

Now：
1、npu_moe_gating_top_k support all size of group_count
2、the functionality of torch_npu.npu_moe_gating_top_k_softmax are included in torch_npu.npu_moe_gating_top_k

CANN: depends on 8.3.RC1

Performance：

1. GLM4.5-w8a8, TPS improve 6%
2. Qwen3, the same as before

Does this PR introduce any user-facing change?

How was this patch tested?

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request adds support for a fused gatingtopk MoE operator. The changes introduce a check function, _check_npu_moe_gating_top_k, to determine if the fused operator can be used, and refactors the expert selection logic to use this check. The tests have also been updated to reflect these changes.

My review has identified a critical bug in _check_npu_moe_gating_top_k (and its duplicated version in the test file) that could lead to a ZeroDivisionError. The order of checks is incorrect, using num_expert_group as a divisor before validating it. Please see the detailed comments for the fix.

[gemini-code-assist]

This function has a potential ZeroDivisionError. The variable num_expert_group is used as a divisor on lines 100 and 104, but it is only checked to be greater than 0 on line 106. If num_expert_group is 0, this will cause a crash. To fix this, the check for num_expert_group > 0 should be performed before any division by it. It would also be better to import and use the _check_npu_moe_gating_top_k function from vllm_ascend.ops.moe.experts_selector to avoid code duplication and ensure consistency.

Suggested change

	if top_k < 1 or \
	top_k > (hidden_states.shape[-1] / (num_expert_group * topk_group)):
	return False
	if topk_group < 1 or topk_group > num_expert_group:
	return False
	if topk_group * hidden_states.shape[-1] / num_expert_group < top_k:
	return False
	if not (num_expert_group > 0 and hidden_states.shape[-1] % num_expert_group
	== 0 and hidden_states.shape[-1] // num_expert_group > 2):
	return False
	if not (num_expert_group > 0 and hidden_states.shape[-1] % num_expert_group
	== 0 and hidden_states.shape[-1] // num_expert_group > 2):
	return False
	if top_k < 1 or \
	top_k > (hidden_states.shape[-1] / (num_expert_group * topk_group)):
	return False
	if topk_group < 1 or topk_group > num_expert_group:
	return False
	if topk_group * hidden_states.shape[-1] / num_expert_group < top_k:
	return False

[gemini-code-assist]

There is a potential ZeroDivisionError in this function. The variable num_expert_group is used as a divisor on lines 115 and 119 before it is checked to be greater than 0 on line 121. If num_expert_group is 0, the code will crash. The check for num_expert_group > 0 should be moved before it is used in any division operations to prevent this.

Suggested change

	if top_k < 1 or \
	top_k > (hidden_states.shape[-1] / (num_expert_group * topk_group)):
	return False
	if topk_group < 1 or topk_group > num_expert_group:
	return False
	if topk_group * hidden_states.shape[-1] / num_expert_group < top_k:
	return False
	if not (num_expert_group > 0 and hidden_states.shape[-1] % num_expert_group
	== 0 and hidden_states.shape[-1] // num_expert_group > 2):
	return False
	if not (num_expert_group > 0 and hidden_states.shape[-1] % num_expert_group
	== 0 and hidden_states.shape[-1] // num_expert_group > 2):
	return False
	if top_k < 1 or \
	top_k > (hidden_states.shape[-1] / (num_expert_group * topk_group)):
	return False
	if topk_group < 1 or topk_group > num_expert_group:
	return False
	if topk_group * hidden_states.shape[-1] / num_expert_group < top_k:
	return False