feat(refit): view-ops slice planning for reshard

[tanushriya910]

What

Extend reshard geometry capture and the slice planner to support rank-reducing view ops, so a trainer that publishes a stacked source (e.g. MoE experts stacked into one tensor) can be resharded into its per-slice destination weights.

Changes

• geometry.py — add select and split to _SUPPORTED_OPS. select (rank-reducing) and split (multi-return view, like chunk) are how a stacked source is unstacked into per-expert / per-slice tensors. Without them, every such weight was classified unsupported and the receiver failed closed.
• slice_plan.py — _view_to_box_perm now derives the source box for rank-reducing views (select/unbind drop a dim): a dropped source dim is not iterated by any view dim, so it contributes a single fixed index ([idx, idx+1]) decoded from the offset. Rank increase (unsqueeze) and dim-merging reshapes still raise UnsupportedReshard.
• Tests: test_reshard_refit_geometry.py, test_reshard_refit_slice_plan.py.

Why

Stacked-expert MoE layouts (and similar stacked sources) need select/split to map each slice onto its own destination box — closing that gap in view-op coverage.

Stacked on the receiver-mapping PR (base tanushriyas/reshard-native-convert); retarget to main once that merges.