We only care if we have changed the mode and we're changing it back to its
original.  Even if the current mode matches the target mode, we may still
need to release the displays and clear the entry from originalDisplayModes.

winemac: Don't assume the current display mode is the original if we don't have the displays captured.

Another process may have changed the display mode before we queried the
current mode, so we may be seeing a non-original mode.

originalDisplayModes should be used when active, empty when inactive.
latentDisplayModes is used when inactive, empty when active.

The count of entries in originalDisplayModes is used to test whether the
process has the displays captured so adding entries when inactive would give
incorrect results.  This could have led us to mistakenly change the display
mode when we don't have the displays captured.

The ARR codepath used if ctx->target_version >= NV2 does not account for
the correction offset, but the indirect addressing reads do, which
ultimately results in incorrect offsets. Since the limitation goes away
as soon as we enable GL_NV_vertex_program_2_option ignoring the offset
is the best thing to do. I don't set shader_data->rel_offset to zero in
the pre-compilation analysis because enablding NV_vertex_program2 is in
theory a decision that is taken later during code generation.

dxdiagn: Don't fail the whole initialization for invalid filter data in build_directshowfilters_tree.