[XPU] refactor moe ffn (#5501)

- remove BKCL_DISPATCH_ALL_GATHER
- support sparse mode
- support moe quant_method

fastdeploy/model_executor/layers/quantization/w4a8.py

@@ -52,6 +52,6 @@ class W4A8Config(QuantConfigBase):
                 XPUW4A8MoEMethod,
             )
 
-            return XPUW4A8MoEMethod(self)
+            return XPUW4A8MoEMethod(self, layer)
         else:
             raise ValueError(f"Unsupported layer type {type(layer)} for w4a8")

fastdeploy/model_executor/layers/quantization/weight_only.py

@@ -101,7 +101,7 @@ class WeightOnlyConfig(QuantConfigBase):
                     XPUWeightOnlyMoEMethod,
                 )
 
-                return XPUWeightOnlyMoEMethod(self)
+                return XPUWeightOnlyMoEMethod(self, layer)
             else:
                 from fastdeploy.model_executor.layers.backends import (
                     XPUWeightOnlyLinearMethod,