Add ROCM_EXPLICIT_CTX support to internode.cu

[avinashkethineedi]

Motivation

The dispatch and combine kernels had a fixed RDMA context selection. This adds flag-driven selection so the same kernels can be built for explicit, work-group, or context-free modes, making it easier to tune performance and debug context issues.

Technical Details

• Replaced the 2-way (ROCM_DISABLE_CTX vs. rocshmem_ctx_array) preprocessor branches in the dispatch and combine kernels with a 3-way selection:
  • ROCM_EXPLICIT_CTX → per-SM explicit context (rocshmem_ctx_array[sm_id])
  • default (neither flag) → shared work-group context (ctx)
  • ROCM_DISABLE_CTX → context-free shmem fallback
• Enabled the work-group ctx create/destroy (shmem_wg_ctx_create / shmem_wg_ctx_destroy) for the default path in both dispatch and combine.
• Guarded the combine ctx destroy with USE_ROCM so it stays symmetric with its create, avoiding an undeclared ctx reference in non-ROCm builds.
• Removed a leftover commented-out std::cout debug block and the now-unused <iostream> include.

Test Plan

• Built with ROCM_EXPLICIT_CTX enabled and ran the internode stress test for ~150 iterations

Test Result

• Successfully passed all 150 iterations with no failures.

Submission Checklist

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

[avbokovoy]

Hi @avinashkethineedi, @ahubbe-amd

Thanks for your contribution. I noticed that there's a possibility for the user of DeepEP library to specify both ROCM_EXPLICIT_CTX and ROCM_DISABLE_CTX at the same time by either passing --disable-ctx and -DROCM_EXPLICIT_CTX flags to setup.py or -DROCM_EXPLICIT_CTX -DROCM_DISABLE_CTX. Could you please confirm if this scenario should be avoided? If so, let's do the following:

1. Add --expicit-ctx flag with detailed hint to setup.py with runtime check that only one of --disable-ctx or -expicit-ctx flags is specified. Or none
2. Add static_assert( defined(ROCM_EXPLICIT_CTX) && (ROCM_DISABLE_CTX) ) into internode.cu file to guard against defining both variables.

Minor: Probably ROCM_EXPLICIT_CTX naming doesn't reflect the purpose of the variable. Might be a good idea to reconsider it.