LoongArch: Avoid using $r0/$r1 as "mask" for csrxchg

LoongArch: Avoid using $r0/$r1 as "mask" for csrxchg

When building kernel with LLVM there are occasionally such errors:

In file included from ./include/linux/spinlock.h:59:
In file included from ./include/linux/irqflags.h:17:
arch/loongarch/include/asm/irqflags.h:38:3: error: must not be $r0 or $r1
   38 |                 "csrxchg %[val], %[mask], %[reg]\n\t"
      |                 ^
<inline asm>:1:16: note: instantiated into assembly here
    1 |         csrxchg $a1, $ra, 0
      |                       ^

To prevent the compiler from allocating $r0 or $r1 for the "mask" of the
csrxchg instruction, the 'q' constraint must be used but Clang < 21 does
not support it. So force to use $t0 in the inline asm, in order to avoid
using $r0/$r1 while keeping the backward compatibility.

Cc: stable@vger.kernel.org
Link: https://github.com/llvm/llvm-project/pull/141037
Reviewed-by: Yanteng Si <si.yanteng@linux.dev>
Suggested-by: WANG Rui <wangrui@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>