Issue is rather well explained in T45471: our current customdata writing code easily generates several different blocks in blend file with same 'old' address. This is bad, because those addresses are used as 'uid' during reading process (it kind of work in Blender's own reading process, by mere luck mostly, but breaks the file specs).

Solution (suggested by Campbell, thanks) implemented by this patch is to avoid duplicating everything, and instead just overwrite what we needs to skip some cdlayers on write:

• the CustomData's totlayer number;
• the CustomData's layers array of CustomDataLayer (keeping its original address using the writestruct_at_address helper).

New design allows us to get completely rid of the no_free flag stuff in write_customdata().

Note that this implies written data is not directly valid from Blend PoV, since its written typemap does not match written layers (this is not an issue because typemap is rebuilt on read anyway - and it's easy to fix this if really needed).

Also, the backward compatibility saving of mface data remains an issue here, see comment in code.