tu: Don't patch GMEM for input attachments never in GMEM

This can happen if we resolve to a resolve attachment and then use that
resolve attachment as an input attachment in a later subpass. We don't
need to put it in GMEM, but it's still considered "written" because
input attachment reads need a dependency after the resolve.

MSRTSS input attachment tests effectively created such a scenario after
lowering to transient multisample attachments and inserting resolves.

Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37919>

src/freedreno/vulkan/tu_pass.cc

@@ -425,7 +425,8 @@ tu_render_pass_patch_input_gmem(struct tu_render_pass *pass)
             written[a] = true;
             subpass->feedback_invalidate = true;
          }
-         subpass->input_attachments[j].patch_input_gmem = written[a];
+         subpass->input_attachments[j].patch_input_gmem =
+            written[a] && pass->attachments[a].gmem;
       }
 
       for (unsigned j = 0; j < subpass->color_count; j++) {