ARM: Use 64-bit literals for LoadDImmediate().

And rewrite the medium-range long/fp literal to use
MOVW+ADD+LDRD/VLDR because the old instruction sequence
was broken if the "ADD ip, pc" was not 4-byte aligned.

Test: assembler_thumb2_test has been updated. Standard
      ART test suite has been run on host and Nexus 5.
Change-Id: I37c6a62aa6e77c6a9701b5a1fb4db2e666c1eae9