libvulkan: Fix double-free, refactor instance destruction

Fixes dEQP-VK.api.object_management.alloc_callback_fail.instance.
Since we were calling DestroyInstance_Bottom from both
CreateInstance_Bottom and CreateInstance_Top failure paths, we were
calling the driver's DestroyInstance twice.

To avoid such bugs, this change clears the driver instance handle to
VK_NULL_HANDLE after calling the driver DestroyInstance.

But the real fix in this change is to make creation and destruction
symmetric. Now DestroyInstance_Bottom only cleans up the things that
were initialized/allocated in CreateInstance_Bottom, and is only
called from CreateInstance_Bottom failure paths and from a dispatched
vkDestroyInstance. Similarly, DestroyInstance_Top and failure paths in
CreateInstance_Top call DestroyInstance (formerly TeardownInstance) to
clean up things initialized/allocated in CreateInstance_Top. The
direct calls from *_Top functions to DestroyInstance_Bottom are gone
-- *_Top functions should only reach *_Bottom functions via dispatch,
so the call goes through enabled layers.

Bug: 27493757
Change-Id: I4e9f8508297813415499dc17803fff49ce9abdcf
(cherry picked from commit 15cd1e269fd2dacef8b95006928b122b9dabbeea)