[None][feat] add Rubin topology support to GVR V2 prefill

[longcheng-nv]

Dependencies

Built on #18446, #18702, and the companion decode Draft #18763. Keep this PR in Draft until those dependencies are settled, then rebase out the duplicated base commits.

Summary

• Carry the self-sampling GVR V2 prefill path from [None][feat] Self-sampling GVR V2 prefill indexer top-K #18702 to SM107.
• Replace B200-specific launch assumptions with runtime SM topology.
• Preserve the existing prefill interval, input-format, short-row, slab, and radix-fallback behavior.
• Retain the inherited default-off decode locality-domain prototype, including its BS=1 full-device fallback and fail-closed capture/device checks.

Current scope

Prefill deliberately remains a full-device launch. The current path has one logical row and one CTA per row, so dispatching it to two domains would add launch/event overhead without creating more row-level parallelism or memory locality.

• The indexer producer and logits remain global/full-device.
• The low-level launch is topology/domain aware, but production prefill does not enable dual-domain routing.
• TMA, PDL, localized logits/K-cache allocation, and producer-side locality are not implemented.
• The inherited decode prototype is limited to one in-flight model execution per CUDA device.
• No R200 correctness or performance result is claimed.

Validation

Passed locally on the restacked branch:

• GVR decode host tests: 11/11.
• Rubin prefill topology tests: 10/10.
• Decode route parity with the companion decode branch: 21,504 cases.
• Decode streaming parity: 43,008 comparisons.
• Prefill R=1 route/topology invariants: 21,504 cases.
• Relevant pre-commit formatting/lint, Python compilation, DCO, and whitespace checks.

Full package/GPU pytest collection is blocked in the current environment by the external bindings.so mismatch (global_steady_clock_now). The installed CuTe DSL also lacks the Rubin helper, so SM107 compile, eager, CUDA Graph, correctness, and performance paths have not been exercised.

Draft exit criteria

Before marking Ready:

• Run the companion decode R200 matrix over every BS=1-1024 and compressed N={4K,8K,16K,32K,64K,128K,256K} on this complete stacked branch.
• Validate prefill correctness for K={512,1024,2048}, supported compression ratios, short rows, unaligned [ks, ke) intervals, long rows, and slab-boundary cases.
• Exercise eager execution and CUDA Graph warmup/replay on strict and balanced Rubin topology.
• Measure prefill and indexer + Top-K end-to-end latency against the full-device baseline.
• Confirm that SM100/SM103 behavior and unsupported-path fallbacks remain unchanged.