This fixes a regression introduced by ebbcc10b.
Prior to that commit, buffers which were not CPU-accessible would not have the
WINED3D_BUFFER_USE_BO flag set, and accordingly
buffer_resource_sub_resource_map() would simply return the SYSMEM location,
ignoring the DISCARD and NOOVERWRITE flags.

However, the "accelerated" path in wined3d_cs_map_upload_bo() only checks for
the DISCARD flag, assuming that it is only set for dynamic GPU-accessible
buffers, and would subsequently try to allocate non-mappable memory and then map
it.

This commit avoids the accelerated path for such buffers, once again matching
our old behaviour. According to [1], this also matches Windows: DISCARD on
SYSTEMMEM buffers is ignored on both AMD and NVidia, and NOOVERWRITE is ignored
on NVidia (but not AMD).

[1] https://www.winehq.org/mailman3/hyperkitty/list/wine-devel@winehq.org/message/XACSAVAGYTJUZDFLX4JWO665VTB6M7LX/

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=53863