fix: preserve 16-bit depth in LDM3D rgblike_to_depthmap

[sergioperezcheco]

Problem

VaeImageProcessorLDM3D.rgblike_to_depthmap combines two 8-bit channels into a 16-bit depth value using widened arithmetic (int32), but then casts the result back to the original uint8 input dtype. This truncates any depth value > 255 to the low byte.

Downstream, postprocess(output_type="pil") calls numpy_to_depth, which constructs a mode="I;16" PIL image expecting 2 bytes per pixel. The truncated uint8 buffer is half the expected size, causing:

ValueError: buffer is not large enough

Fix

Both the torch.Tensor and np.ndarray branches now return uint16 instead of casting back to the input dtype. This preserves the full 16-bit depth range [0, 65535] and is consistent with what numpy_to_depth expects.

Removed the now-unused original_dtype variable in both branches.

Test

Added test_rgblike_to_depthmap_preserves_uint16_range verifying:

• NumPy: uint8 input with high byte=1, low byte=0 → uint16 output with value 256
• Torch: same check

Fixes #14206

[sergioperezcheco]

Closing this as a duplicate of #14200 by @chuenchen309, which was opened ~30h earlier and addresses the same bug from #14206. Both PRs converge on uint16 for the numpy path; the only difference is mine also casts the torch path to uint16 whereas #14200 keeps int32 there. Happy to leave that dtype question to the maintainers rather than fragment review across two PRs. Thanks!