For large documents we create and destroy a large number of non-poolable
SfxPoolItems, which get inserted into and removed from a vector.
Unfortunately the performance of this (depending on pattern) is O(N) and
this insert/remove/extend pattern can happen per text span we insert.
This patch makes this O(const) via a hash. This gives a 5x speedup for
the above bug; 176s to 34s or so, and moves the remaining performance
issues elsewhere.

Unfortunately, we have to retain the ordered array to keep the binary
file format code (used for editeng cut-and-paste) in place, so have to
keep both a hash, and an array, and a list around for free slots. cf.
fdo#79851 where there is a start at removing that.

This wastes space; but not that much - for a large open document
collection we have O(100's) of SfxItemPools, and O(1000's) of
SfxPoolItemArray_Impls; having fixed fdo#79851 we can consolidate this.

Add skeletal unit test; translate several German comments; remove
un-necessary include.

Change-Id: Ie0de32b1a29217560c5591c71a6cd4e26d39a531