LoongArch: Relax memory ordering for atomic operations

This patch relaxes the implementation while satisfying the memory ordering
requirements for atomic operations, which will help improve performance on
LA664+.

Unixbench with full threads (8)
                                           before       after
  Dhrystone 2 using register variables   203910714.2  203909539.8   0.00%
  Double-Precision Whetstone                 37930.9        37931   0.00%
  Execl Throughput                           29431.5      29545.8   0.39%
  File Copy 1024 bufsize 2000 maxblocks    6645759.5      6676320   0.46%
  File Copy 256 bufsize 500 maxblocks      2138772.4    2144182.4   0.25%
  File Copy 4096 bufsize 8000 maxblocks   11640698.4     11602703  -0.33%
  Pipe Throughput                          8849077.7    8917009.4   0.77%
  Pipe-based Context Switching             1255108.5    1287277.3   2.56%
  Process Creation                           50825.9      50442.1  -0.76%
  Shell Scripts (1 concurrent)               25795.8      25942.3   0.57%
  Shell Scripts (8 concurrent)                3812.6       3835.2   0.59%
  System Call Overhead                     9248212.6    9353348.6   1.14%
                                                                  =======
  System Benchmarks Index Score               8076.6       8114.4   0.47%

Signed-off-by: WANG Rui <wangrui@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>