perf(pagerank): add fast, bounded, and memory-aware auto paths

[lmeyerov]

Stack

Draft follow-up stacked on #2006. Review only the four commits in this PR; merge #2006 first.

Summary

• add explicit fast, bounded, and conservative memory-aware auto execution modes to graphistry.std.pagerank
• keep the cuGraph-compatible signature/defaults plus Graphistry extras, direct/GFQL parity, original IDs, weighted/personalized/nstart/precomputed behavior, and shared convergence semantics
• preserve the dataframe chunked path for bounded memory and graph500-26-scale workloads
• use persistent compact NumPy/CuPy buffers for fit-in-memory execution; cuDF uses direct atomic scatter and fused rank update
• never silently delegate std execution to igraph or cuGraph

Selection policy

auto is the default. With chunks=1, it selects fast only when a conservative calibrated scratch estimate is at most half detected free host/device memory. Unknown or tighter memory falls back to bounded. Explicit fast bypasses the estimate; explicit bounded preserves the chunkable dataframe path.

Matched evidence

Directed unweighted cit-Patents, prepared solver calls, 3 samples / 1 warmup:

backend	median	peak memory evidence
final std-pandas fast	1.386604s	372.05 MiB
igraph PRPACK	1.449537s	baseline
final std-cuDF auto/fast	0.164078s	262.793 MiB RMM delta
cuGraph 26.02	0.035157s	baseline

Final GPU samples: [0.166276, 0.164078, 0.162755]; std-cuDF is 4.67x cuGraph and clears the <=10x target. Rank mass is 1.0000000000000013, maximum retained delta is 8.131516293641283e-20, and top-10 IDs match. The prior std-cuDF baseline was 1.783722s / 979.153 MiB. The bounded graph500-26 memory result was preserved and cuGraph graph500-26 was not rerun.

Artifact SHA-256: 91b44ed81cb4b5b43beee6b341c2d8ce6369d4cedc0e7bde9f821adfc13ff53f.

The final DGX runtime artifact was built from implementation commit 7022fe742573e4212339bb326ad887f29b03f518; current PR head ec039cdea784036229d66d43d7752751a1db1671 adds CPU-only coverage contracts and does not change runtime code.

Validation

• scoped lint/hygiene: pass
• full configured mypy 2.3.1: pass, 337 source files
• local focused CPU: 62 passed, 1 explicitly gated cuDF skip
• exact CI-baseline changed-line coverage reproduction plus focused tests: 81.15% versus 80% gate
• guarded DGX semantic lane with TEST_CUDF=1: 56 passed, no skips
• guarded DGX matched performance/correctness/memory lane: pass

All Docker, GPU, dataset, profiling, and performance work ran only on dgx-spark through the perf lock and safe_run.sh, with one no-network container, RMM/cgroup caps, >=80 GiB host-memory floor, watchdog, and hard timeout. Companion benchmark PR: https://github.com/graphistry/pyg-bench/pull/194

[lmeyerov]

Thought should be clear that the main value of this implementation is lower memory consumption for handling bigger graphs on smaller GPUs

[lmeyerov]

Likewise, typically better to use the GPU or CPU ones when fit in memory

[lmeyerov]

Send most of these live in our SeriesT / DataframeT cross platform files ?

[lmeyerov]

pagerank code should be in its own pagerank.py, same for any other algorithms