yes: strange performance gap between yes vs yes no (3 bytes is much faster than 2 bytes)

`yes` has multiple of 2 bytes output (y\newline, tee() only code path)
while `yes no` has multiple of 3 bytes (tee() + splice() code path).
This happens only with main branch's `pv`.
```
$ yes y | pv>/dev/null
35.5GiB 0:00:01 [35.5GiB/s]
$ yes no | pv>/dev/null
56.3GiB 0:00:01 [56.3GiB/s]
```
Logically, tee() only should have hald syscalls/sec and should be slightly faster.