modify_ldt always allocates a 64 kB vmalloc block, regardless of how
many ldt entries are actually needed.
All apps linked to libpthread now use the ldt to access thread-local
variables, and an 8192 entry ldt is both a waste of memory and of the
vmalloc area.

The attached patch:
* dynamically allocs and reallocs the ldt, starting with a 512 entry (4
kB) ldt allocated with kmalloc.
* returns all 5 entries of the default ldt if modify_ldt(2,,) is called.
* moves copy_segments into init_new_context, and adds error handling.
* deletes copy_segments on i386.
* removes the call to copy_segments from kernel/fork.c. No other
platform uses it.
* reuses cpu_vm_valid for the LDT reload: there is no need to maintain 2
cpu bitmasks, one set_bit removed from every thread switch.

Tested on i386 SMP, both with libpthread apps and a small test app.

--
	Manfred