Since adding MO_ATOM_MASK, the maximum MemOpIdx requires 15 bits,
which overflows the 12 bit field allocated for TCI memory ops.
Expand the field to 16 bits for 2-operand memory ops, and place
the value in TCG_REG_TMP for 3-operand memory ops (same as we
already do for 4-operand memory ops).

Cures a debug assert for aarch64, with FEAT_LSE2 enabled.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>