[BugFix] mega moe max token default value fix (vllm-project#11584)

### What this PR does / why we need it?
Change the default value of mega_moe_max_tokens from 65536 to 131072.

### Does this PR introduce _any_ user-facing change?
NA

### How was this patch tested?
Configurable parameter, no extra testing needed.

- vLLM version: v0.23.0
- vLLM main:
vllm-project/vllm@1f486d9

Signed-off-by: justice-dance <justice1717@163.com>

Author: justice-dance

vllm_ascend/ascend_config.py

@@ -294,8 +294,8 @@ def __init__(self, vllm_config: "VllmConfig"):
         # are dropped and skipped from computation, degrading accuracy.
         # Do not set this too large: workspace memory scales linearly with this value, which matters
         # especially under long-context scenarios where the operator should not hold too much memory.
-        # Default 65536.
-        self.mega_moe_max_tokens = additional_config.get("mega_moe_max_tokens", 65536)
+        # Default 131072.
+        self.mega_moe_max_tokens = additional_config.get("mega_moe_max_tokens", 131072)
         if not isinstance(self.mega_moe_max_tokens, int):
             raise ValueError(
                 f"mega_moe_max_tokens must be an integer, got {type(self.mega_moe_max_tokens).__name__}: "