[lmi] A small slowdown in xml serialization

[Greg Chicares]

Revision 4841 of 20100417T1245Z made reads a little slower, according to
the following 'input_test' results. I believe I can remove the penalty,
but I want to defer that for now. Note that this report concerns (my)
'xml_serializable.?pp' only, and not (Vaclav's) 'xml_serialize.hpp'.

revision 4840:
  Overhead: 3.905e-06 s =       3905 ns, mean of 2561 iterations
  Read    : 4.251e-03 s =    4250971 ns, mean of 100 iterations
  Write   : 3.015e-03 s =    3014961 ns, mean of 100 iterations
  'cns' io: 4.408e-02 s =   44083512 ns, mean of  23 iterations
  'ill' io: 1.090e-02 s =   10903121 ns, mean of  92 iterations

revision 4841:
  Overhead: 4.023e-06 s =       4023 ns, mean of 2486 iterations
  Read    : 4.967e-03 s =    4967119 ns, mean of 100 iterations  +17%
  Write   : 2.970e-03 s =    2969952 ns, mean of 100 iterations  - 1%
  'cns' io: 4.659e-02 s =   46593767 ns, mean of  22 iterations  + 6%
  'ill' io: 1.168e-02 s =   11684165 ns, mean of  86 iterations  + 7%

xml_serializable<T>::{read,write}_element() could have been written
in terms of xml_serialize_hpp::{get,set}_element(), but that costs:
  Read    : 5.140e-03 s =    5140403 ns, mean of 100 iterations  +21%
  Write   : 4.199e-03 s =    4199027 ns, mean of 100 iterations  +39%
  'cns' io: 5.064e-02 s =   50637585 ns, mean of  20 iterations  +15%
  'ill' io: 1.296e-02 s =   12956410 ns, mean of  78 iterations  +19%
because of the overhead that the more general approach demands.

The speed of a complete system test hasn't changed materially:
  r4840 make system_test > ../log 2>&1  82.96s user 68.52s system 44% cpu 
5:36.93 total
  r4841 make system_test > ../log 2>&1  83.08s user 66.67s system 44% cpu 
5:37.89 total
Here, I'm running an msw native binary in Cygwin zsh, so the output of
'time' isn't necessarily meaningful except for the 'total'.