To be used in regression testing and similar scenarios, where the output ODF
is *not* intended to be further manipulated in LibreOffice. An environment
variable LIBO_ONEWAY_STABLE_ODF_EXPORT is used to toggle this behaviour. I am
not 100% sure whether the generated ODF with the hack toggled on is even fully
correct, but correctness is not the purpose of the hack anyway.

Two classes of issues handled: 1) Automatic style names and 2) use of
randomness.

For class 1), when the hack toggle is in effect, we generate the names at
first as strings based on all the properties of the style, and sort them based
on those, and then rename them (for brevity in the output) to the "normal"
form of a short prefix plus a number (like "P12").

Sure, it would have been better to just figure out *why* the automatic style
naming currently is not stable in the first place, but outputs the styles in
different order (with some styles being assigned different numbers) in
separate invokations of LibreOffice), but I was unable to understand that.

Possibly this code could be used in all cases, except that it does break some
unit test (can't recall which right now). I don't know whether that is simply
because the unit test assumes too much knowledge of the internal workings of
the automatic style name generation, or whether the generated ODF is actually
invalid.

For 2), I found a handful of places where randomness was used to generated
various kinds of identifiers in ODF output. I changed those to just use large
(64-bit) non-overlapping integers instead. I assume there *is* a point in the
original code in each case that explains why randomness is needed, so the hack
definitely needs to be optional and used only for the above mentioned
scenarios.

Change-Id: I17b657197e38bcf24abdfe61ad4a277f4339eeae