huge performance impact when order by PK prefix

Version: 22.3.8.39-lts

How to reproduce:
```sql
CREATE TABLE login_log
(
`app_id` UInt8,
`company_id` UInt32,
`login_date` Date CODEC(DoubleDelta, LZ4),
`login_hour` UInt8,
`user_name` String CODEC(ZSTD(1)),
INDEX idx_sid company_id TYPE bloom_filter GRANULARITY 4
)
ENGINE = MergeTree
PARTITION BY toYYYYMM(login_date)
ORDER BY (app_id, login_date)
SETTINGS index_granularity = 8192;

sql1:
SELECT company_id as userId,
login_date as loginDate,
login_hour as loginHour,
app_id as appId,
user_name
from user_login_log prewhere company_id= 1
where app_id= 5
and login_date< '2022-10-01'
ORDER by app_id DESC,
login_date DESC,
login_hour DESC
limit 1;
-- 1 rows in set. Elapsed: 5.847 sec. Processed 1.10 billion rows, 25.15 GB (188.53 million rows/s., 4.30 GB/s.)

sql2:
SELECT company_id as userId,
login_date as loginDate,
login_hour as loginHour,
app_id as appId,
user_name
from user_login_log prewhere company_id= 1
where app_id= 5
and login_date< '2022-10-01'
ORDER by login_date DESC,
login_hour DESC
limit 1;
-- 1 rows in set. Elapsed: 0.224 sec. Processed 3.75 million rows, 81.17 MB (16.76 million rows/s., 362.59 MB/s.)
```

the difference between sql1 and sql2 is only: sql1 have ORDER by app_id DESC, sql2 do not have. but sql1 scanned 25.15 GB data, sql2 only scanned 81.17 MB data.

what's the reason of the performance difference above?