The bee inventory file `/var/cache/bee/beecache/INVENTORY` contains the
filenames of the files installed from bee.

One problem is, that a file can have more than one name name because of
symlinks. For example, we currently have a symlink `/usr/doc ->
share/doc` and the inventory currently contains this entry:

   cunit-2.1_3-1.x86_64 1619448799 0 0 0100644 683 7eb1686ef4ebac0f4743c407da5aab52 /usr/doc/CUnit/CUnit_doc.css

So the file is not registerd by its canonical name
`/usr/share/doc/CUnit/CUnit_doc.css` but by the alias name
`/usr/doc/CUnit/CUnit_doc.css`. This can happen, if a package is
installed "through" a symlink in the system.

The problem with multiple paths to the same file is that if one version
of a package installs a file through one path and another version of a
package installs the same file though another path, then the file is
lost after "bee update".

This is because all files from the old package, which are not
registered by the new version of the same package (or any other
installed package) are removed. While the removal works through a
symlink, the protection by its single registered filename does not.

To mitigate this problem, bee should protect the file itself during
certain operations like `bee update`. It is not enough to protect
name variant used in the inventory.

We want to achieve this by translating the directory names used in the
inventory to their canonical form (if possible).

This patch adds a filter tool for the bee index format, which translates
the filenames into a canonical form.

Don't just use realpath(3) or canonicalize_file_name(3) for that,
because this would be much to slow. These functions do readlink() for
every path component of the provided name. As we have to do this for
every file of the inventory, the system call usage and file system
access would explode. Processing the inventory file this way, took more
than two minutes.

Instead, cache the results of readlink for a single invocation of the
tool, so that the operation is done only once per file. Use another
cache for complete translated path names to reduce the load to the
readlink-cache. This way, the same result can be achived in 0.3 seconds.

    $ ls -ld /usr/doc
    lrwxrwxrwx 1 root root 9 Mar  7  2011 /usr/doc -> share/doc
    $ grep CUnit_doc.css /var/cache/bee/bee-cache/INVENTORY
    cunit-2.1_3-1.x86_64 1619448799 0 0 0100644 683 7eb1686ef4ebac0f4743c407da5aab52 /usr/doc/CUnit/CUnit_doc.css
    $ ./beeindextr /var/cache/bee/bee-cache/INVENTORY | grep CUnit_doc.css
    cunit-2.1_3-1.x86_64 1619448799 0 0 0100644 683 7eb1686ef4ebac0f4743c407da5aab52 /usr/share/doc/CUnit/CUnit_doc.css

When "bee query PATTERN" is used to grep for installed files, use
the inventory file and translate the filenames from the inventory to
their canonical form, which is the real path, not the path used to
install them, possibly following symlinks.

The new code is a lot faster than the old one. The wall time for
`fakeroot bee query otop` is reduced from 0m11.712s to 0m0.760s.