fix: batch collapse, memory blowup, resize rounding, and 8 more Should-Fix issues from code review

- Reject use_original_name=False + use_counter=False (all files collapsed to one output name)
- Add preview_limit to cap in-memory preview tensors (was holding entire batch as float32)
- Recover from per-file failures instead of aborting the whole batch
- Fix _calc_new_size overshoot past longest_side and off-by-one truncation
- Add upscale_if_smaller to SmartBatchResize to match SmartResize
- Clip pixel values before uint8 cast (was wrapping on overflow)
- Support CUDA tensors and RGBA/grayscale images in SmartResize
- Handle empty batches; derive width/height from the actual output tensor
- Require POST + loopback/origin check on the browse_folder endpoint (CSRF)
- Add subprocess timeout to the folder picker dialog
- Isolate the browse_folder route registration so import failures don't take down all nodes
- Validate folder_custom against path traversal and self-overwrite
- Replace IS_CHANGED=NaN with a folder-listing hash so unrelated graph runs can still cache

example_workflows/smart_batch_resize_demo.json

@@ -33,6 +33,7 @@
         1024,
         8,
         "lanczos",
+        true,
         "",
         true,
         false,
@@ -41,7 +42,8 @@
         false,
         "resized",
         true,
-        "_"
+        "_",
+        32
       ]
     },
     {