Fixes #20439 .

I am unsure why transfer queues add such a huge overhead(it was at least 10GB as tried reserving that much when loading the model and it still choked)

Summary from big C:

Using the separate SDMA/transfer queue for async copies is counterproductive here because:

1. There's no separate device memory to transfer to/from — it's all unified memory
2. The SDMA engine and its associated driver structures (kernel buffer objects for command streams, page table entries, cross-queue synchronization state) consume memory from the same pool
3. The timeline semaphore synchronization between compute and transfer queues adds driver overhead with no benefit — the compute queue can issue buffer copies just as efficiently on UMA
4. The transfer queue's command pool (transfer_cmd_pool) and its command buffers accumulate during model loading alongside compute_cmd_pool, effectively doubling the command infrastructure