Use upstream DTensor select scatter strategy

[aditvenk]

Stacked PRs:

• Use upstream DTensor expand strategy #504
• Use upstream DTensor reshape strategy #503
• Use upstream DTensor scatter strategy #502
• Use upstream DTensor convolution backward strategy #501
• Use upstream DTensor convolution strategy #500
• Use upstream DTensor layer norm backward strategy #499
• Use upstream DTensor layer norm strategy #498
• ->Use upstream DTensor select scatter strategy #497
• Use upstream DTensor diagonal scatter strategy #496
• Use upstream DTensor convert strategy #495
• Use upstream DTensor constant pad strategy #494
• Use upstream DTensor index strategy #493
• Use upstream DTensor transpose strategy #492
• Use upstream DTensor reduction strategy #491
• Make AC recompute test topology-robust #490

---

Use upstream DTensor select scatter strategy

PyTorch now provides a single-dim strategy for aten.select_scatter.default with the same non-selected-dimension sharding mapping, so AutoParallel can remove its local rule.

Authored with Claude.

[fmassa]

LGTM, thanks!

[aditvenk]

closing in favor of #505