feat: expand OpenCL operator support (where, channel_shuffle, ceil_mode)

[cafeTechne]

This PR expands the OpenCL backend with several native operator implementations to reduce CPU fallbacks and improve model compatibility:

• Implemented aten::where: Uses a native ternary broadcasted pointwise kernel.
• Implemented aten::channel_shuffle: Native decomposition (view/transpose/reshape) ensuring data stays on GPU.
• Enabled pooling ceil_mode: Verified and unlocked forward/backward for avg/max pooling.
• Fixed masked_select broadcasting: Added expansion logic to support arbitrary mask shapes.
• Verification: All changes verified with \ est_op.py\ on AMD RX 6500 XT.

Includes submodule updates for dlprimitives fixing kernel parameter syntax.