[refactor] refactor weight trans nz and transpose (#4878)

### What this PR does / why we need it?

Now `VLLM_ASCEND_ENABLE_NZ` will have three options:
0: disable nz;
1: only quant case enable nz;
2: enable nz as long as possible;

And `VLLM_ASCEND_ENABLE_NZ`=1 by default.

All cases are shown in the table below:

|  | W4A4 | W4A8 | W8A8 | fp16/bf16 | fp32 |
|---|---|---|---|---|---|
| trans nz | can't support nz | trans nz by default | trans nz by
default | trans nz when VLLM_ASCEND_ENABLE_NZ is 2 | can't support nz |
| transpose | only support not transpose case | only support transpose
case | only support transpose case | linear: only support not transpose
case<br>gmm: only support transpose case | same to fp16/bf16 |

Some exceptional cases:
1. MLAPO op need to do some additional processing on the weights,
including trans nz. If use MLAPO op, some weight will be transformed to
nz forcely;
2. MLA/SFA's weight `W_UV` will be used by op
`torch.ops._C_ascend.batch_matmul_transpose`, and this op can't support
nz currently;

### Does this PR introduce _any_ user-facing change?
Now fp16/bf16 weight will not trans nz by default.

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: zzzzwwjj <1183291235@qq.com>

vllm_ascend/ops/linear.py

@@ -24,7 +24,6 @@ from typing import Optional, Union
 
 import torch
 import torch.nn as nn
-import torch_npu
 from torch.nn.parameter import Parameter
 from vllm.config import get_current_vllm_config
 from vllm.distributed import divide
@@ -37,7 +36,7 @@ from vllm.model_executor.layers.quantization.base_config import \
 from vllm.model_executor.utils import set_weight_attrs
 
 from vllm_ascend.ops.linear_op import get_parallel_op, get_replicated_op
-from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, is_enable_nz
+from vllm_ascend.utils import maybe_trans_nz
 
 
 class AscendUnquantizedLinearMethod(UnquantizedLinearMethod):
@@ -45,11 +44,8 @@ class AscendUnquantizedLinearMethod(UnquantizedLinearMethod):
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         super().process_weights_after_loading(layer)
-        if "conv1d" not in layer.prefix and (
-                is_enable_nz() and layer.weight.data.dtype
-                in [torch.float16, torch.bfloat16]):
-            layer.weight.data = torch_npu.npu_format_cast(
-                layer.weight.data, ACL_FORMAT_FRACTAL_NZ)
+        if "conv1d" not in layer.prefix:
+            layer.weight.data = maybe_trans_nz(layer.weight.data)
 
 
 # TODO(realliujiaxu): Remove this class after linear of vllm supports custom comm group