[inductor] Save producer partials for downstream sums

[eellison]

Submitted by my agent.

This is a default-off prototype for better-benchmark issue pytorch#21, narrowed to the simple landable shape: existing mixed-order producers can attach compatible downstream producer-body sum consumers as extra saved partials. It does not relax normal mixed-order producer/consumer fusion and does not include the marginal simple-producer path.

Mechanically: while the mixed-order producer materializes a large fp32 full output that is still needed, it also writes compact saved sum partials for values already computed in that producer body. The wrapper then final-reduces the compact workspace instead of rereading the full materialized output/body values.

Current head: c774b530fc8.

Validation:

• Static checks on clean origin/main worktree: py_compile, git diff --check, ruff check.
• Focused GPU tests on built reviewable tree: test/inductor/test_producer_sum_reduction_accumulation.py -v, 4 tests passed.
• 30-candidate B200 sweep artifact: better-benchmark/scratch_issue21_guarded_validation_simplified_64m_20260529.

Why the 64MiB gate:

• A 30-candidate sweep with the earlier 32MiB gate selected one real small regression: sum_sum_sum_69414585b76b went 3 -> 1 kernels but 0.085ms -> 0.090ms.
• Raising the default-off min-size gate to 64MiB rejects that 50MiB case and the old 32MiB boundary case.

Key 64MiB sweep results on B200:

• sum_sum_sum_6107a2f54029: 0.8357ms -> 0.1793ms, kernels 4 -> 1.
• sum_sum_sum_dc96c4651516: 0.8098ms -> 0.1766ms, kernels 4 -> 1.
• sum_sum_sum_e2d4961f3571: 0.0788ms -> 0.0739ms, kernels 3 -> 1.
• sum_sum_sum_59e022113eeb: 0.1115ms -> 0.1066ms, kernels 3 -> 1.
• sum_sum_sum_97999676281e: 0.0933ms -> 0.0907ms, kernels 3 -> 1.
• sum_sum_sum_64f701d26f0a: 0.1725ms -> 0.1679ms, kernels 6 -> 4.
• 30/30 correct, 6 selected, no selected regressions.

This should stay default-off until the profitability model is stronger.