talloc: use a struct for pool headers.
This neatens the code a bit (we should do a similar thing for all the
TALLOC_CHUNK macros).
Two subtler changes:
(1) As a result of the struct, we actually pack object_count into the
talloc header on 32-bit platforms (since the header is 40 bytes, but
needs to be 16-byte aligned).
(2) I avoid VALGRIND_MAKE_MEM_UNDEFINED on memmove when we resize the
only entry in a pool; that's done later anyway.
With -O2 on my 11.04 Ubuntu 32-bit x86 laptop, the talloc_pool speed as
measured by testsuite.c actually increases 10%.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>