I find that in RISC-V, ra is caller saved, in MIPS, ra is callee, which means in RISC-V callee can directly change the value in ra without save, but since ra has changed, how callee return back to caller?
The usage of RISC V ra and MIPS $ra is effectively the same regardless of the designation.
Since both caller (who needs to return to their caller) and (a non-leaf) callee need to repurpose the return address register, the value in that register needs to be preserved. The only logical way to do that is to preserve the register once on entry and restore it once on exit just like the s/$s preserved, callee-saves registers.
However, once thus saved, the return address register may be repurposed by functions for other uses and any such usage would follow caller saves conventions (unlike the $s registers, which are guaranteed to be preserved across a call).
So, effectively, ra/$ra can behave, at different times, both as callee saves, and caller saves.
A caller cannot rely on a value placed into ra/$ra surviving a function call (as they could with $s registers), thus is is caller saves. Yet, when a callee preserves ra/$ra, it preserves it just like it does the $s callee-saves registers — namely in prologue/epilogue.
By contrast, $t registers, if preserved by the caller so as to survive a function call, would have to be preserved after each update to the value (e.g. minimally after the first initialization), and this is caller saves behavior. These registers are initialized first, then preserved, whereas $s registers are preserved first, then initialized.
ra/$ra has behaviors of both callee and caller saves: it needs to be preserved before being initialized and reused/repurposed, which is a callee saves approach, yet, a variable placed into $ra would not survive a function call, and so to survive a function call, would need to be initialized then preserved.