[SQLite] Transaction field handling: intial implementation

[elpaso]

Fix out of sync (not restored) fields after a ROLLBACK.

See: qgis/QGIS#59797 (comment)

This is an SQLite only fix, if the approach sounds reasonable we should consider supporting more drivers.

Funded by: QGIS.org

[rouault]

There's a bit of dissymetry with the call to AddFieldDefn() in CreateField() having been done before the "if IsInTransation())" logic and here where AddGeomFieldDefn() is done after the IsInTransaction() test (note that if you symmetrize things you need to add/remove a -1 index to GetFieldCount/GetGeomFieldCount)

[elpaso]

the reason for the dissimetry is because poGeomField is (already was) a unique ptr whic would be null after the move. I will add a comment.

[elpaso]

comment added

[rouault]

In the documentation of RollbackTransaction() (in ogr/ogrsf_frmts/ogrsf_frmts.dox), it should probably be mentioned that OGRFeature* instances acquired or created between the StartTransaction() and RollbackTransaction() should be destroyed before RollbackTransaction() if the field structure has been modified during the transaction.
In particular, the following is invalid:

lyr->StartTransaction();
lyr->DeleteField(...);
f = new OGRFeature(lyr->GetLayerDefn());
lyr->RollbackTransaction();
// f is in a inconsistent state at that point, given its array of fields doesn't match the updated layer definition, and thus it cannot even be safely deleted !

this must be corrected as:

lyr->StartTransaction();
lyr->DeleteField(...);
f = new OGRFeature(lyr->GetLayerDefn());
delete f;
lyr->RollbackTransaction();

[rouault]

It appears there must be a #define DELETE somewhere in a Windows system header... So maybe switch to Add/Alter/Delete or add a _FIELD suffix ?

[coveralls]

coverage: 70.084% (+0.009%) from 70.075%
when pulling 5ab8543 on elpaso:gpkg-transaction-fields-rollback
into eb84213 on OSGeo:master.

[rouault]

macos_build CI failure unrelated and will be addressed per #11621