More useless inclusions of khash.h

Fix reading attributes of directories

When reading an attribute of a directory there was no guarantee that the
underlying package would be open. When it wasn't reading an attribute
would fail, unless the attribute data were already cached. The reasons
for this are:
* UnpackingDirectory didn't forward the {Init,Uninit}VFS() calls to the
underlying PackageDirectory.
* Only PackageFile was actually opening the package in InitVFS().

Now we forward the {Init,Uninit}VFS() calls in all cases -- even in
{Add,Remove}PackageNode(), when the active package node changes -- and
opening/closing the package is now done in
PackageNode::{Init,Uninit}VFS().

packagefs: Share string instances globally

Introduce a class String which refers to shared private data that is
registered in a global hash table (in class StringPool) and use the
class consequently. This eliminates duplicate allocations for the same
string and also speeds up tests for equality. There's quite a bit
overhead for the management structures (apparently even more than for
the string data itself), but due to almost all strings being used
multiple times this still almost halves the memory usage for string
data.

packagefs: Organize sources into subdirectories

More useless inclusions of khash.h

Fix reading attributes of directories

When reading an attribute of a directory there was no guarantee that the
underlying package would be open. When it wasn't reading an attribute
would fail, unless the attribute data were already cached. The reasons
for this are:
* UnpackingDirectory didn't forward the {Init,Uninit}VFS() calls to the
underlying PackageDirectory.
* Only PackageFile was actually opening the package in InitVFS().

Now we forward the {Init,Uninit}VFS() calls in all cases -- even in
{Add,Remove}PackageNode(), when the active package node changes -- and
opening/closing the package is now done in
PackageNode::{Init,Uninit}VFS().

packagefs: Share string instances globally

Introduce a class String which refers to shared private data that is
registered in a global hash table (in class StringPool) and use the
class consequently. This eliminates duplicate allocations for the same
string and also speeds up tests for equality. There's quite a bit
overhead for the management structures (apparently even more than for
the string data itself), but due to almost all strings being used
multiple times this still almost halves the memory usage for string
data.

packagefs: Organize sources into subdirectories