mips: Fix sub-word atomics implementation

[jrtc27]

Summary

These aligned the address but then always used the least significant
bits of the value in memory, which is the wrong half 50% of the time for
16-bit atomics and the wrong quarter 75% of the time for 8-bit atomics.
These bugs were all present in r178172, the commit that added the mips
port, and have remained for its entire existence to date.