[Bug] Incorrect synchronization insertion in `auto_insert_sync` pass

[fengz72]

**Problem:**The auto_insert_sync pass generates incorrect synchronization logic.

Root Cause: In ascend_sync_insert.cc, the granularity of the inserted PIPE_ALL barrier is too coarse (over-conservative).

Details: The current implementation inserts a PIPE_ALL barrier for every BufferStoreNode, which leads to excessive synchronization and performance degradation.

Involved Code:

void VisitStmt_(const BufferStoreNode* op) override {
    result_.push_back(Evaluate(Call(DataType::Handle(), Op::Get("tl.ascend_auto_barrier"), {StringImm("PIPE_ALL")})));
    result_.push_back(GetRef<Stmt>(op));
}