[Feature] RPC transport splits multimodal tensors incorrectly for multi-image samples

[Wangxiaoxiaoa]

When a sample contains multiple images, multimodal tensors like pixel_values and image_grid_thw may be treated as normal batched tensors by the RTensor/RPC transport layer.

Their leading dimension does not match the outer batch layout, so they can be split or inferred incorrectly during RPC round-trip.

• these multimodal payload tensors should be transportable as non-batched objects
• they should not participate in outer batch layout inference
• they should remain intact if worker methods return mutated batch structures

[garrett4wade]

FYI I've commented in #1038