Fix an assertion in the address space manager. BZ #345887.

The VG_(extend_stack) call needs to be properly guarded because the
passed-in address is not necessarily part of an extensible stack
segment. And an extensible stack segment is the only thing that
function should have to deal with.
Previously, the function VG_(am_addr_is_in_extensible_client_stack)
was introduced to guard VG_(extend_stack) but it was not added in all
places it should have been.

Also, extending the client stack during signal delivery (in sigframe-common.c)
was simply calling VG_(extend_stack) hoping it would do the right thing.
But that was not always the case. The new testcase 
none/tests/linux/pthread-stack.c exercises this (3.10.1 errors out on it).

Renamed ML_(sf_extend_stack) to ML_(sf_maybe_extend_stack) and add
proper guard logic for VG_(extend_stack).

Testcases none/tests/{amd64|x86}-linux/bug345887.c by Ivo Raisr.


git-svn-id: svn://svn.valgrind.org/valgrind/trunk@15138

none/tests/linux/pthread-stack.vgtest

@@ -0,0 +1,2 @@
+prog: pthread-stack
+vgopts: -q