IB/mthca: Avoid alignment traps when writing doorbells

Architectures such as ia64 see alignment traps when doing a 64-bit
read from __be32 doorbell[2] arrays to do doorbell writes in
mthca_write64().  Fix this by just passing the two halves of the
doorbell value into mthca_write64().  This actually improves the
generated code by allowing the compiler to see what's going on better.

Signed-off-by: Roland Dreier <rolandd@cisco.com>