Make copy engine initialization optional

[efric]

Motivation

Iris currently initializes SDMA and creates its host/device queues eagerly when constructing a context, even for workloads that never use SDMA. The issue is that each queue reserves 8 MiB of host/device side memory, with the total allocation growing with the number of local ranks.

Make SDMA opt-in through a new copy_engine parameter on Iris and iris.iris(), defaulting to None. The default skips SDMA initialization, queue creation, and queue-context tensor allocation. SDMA users must explicitly pass copy_engine="sdma" to enable the existing host- and device-initiated operations.

When disabled, get_copy_engine_ctx() returns None, and host SDMA methods raise an error explaining how to enable SDMA. Update the SDMA examples and tests to opt in, and document the new behavior.

Validation: Ruff lint and formatting passed. Targeted tests on two AMD GPUs covered context initialization, symmetric allocation, barriers, SDMA operations, and message passing: 231 passed and 36 existing ROCm graph-capture skips per rank.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[efric]

subsumed by #568