Here we optimize Dsubu by instead of loading imm and subtracting, we
load -imm and perform addition when loading -imm takes less instructions
than loading imm. Similarily li is optimized by loading -imm and
performing addition or loading ~imm and inverting bits using nor when
one of these loads takes two instructions less than loading imm, saving
at least one instruction. Tests are adjusted to cover these
optimizations.

BUG=
TEST=cctest/test-assembler-mips/li_macro
     cctest/test-assembler-mips/Dsubu

Review-Url: https://codereview.chromium.org/2909913002
Cr-Commit-Position: refs/heads/master@{#46001}