Add fused_overlap and multi_overlap dask graph utilities

[brendancol]

Summary

• Adds fused_overlap(agg, *stages, boundary='nan') -- fuses sequential map_overlap calls into a single pass with summed depth, producing one blockwise graph layer instead of N
• Adds multi_overlap(agg, func, n_outputs, depth, boundary) -- runs a multi-output kernel via da.overlap.overlap() + da.map_blocks() in a single pass instead of N separate calls + stack
• Both available as standalone imports and on the .xrs DataArray accessor
• Includes _normalize_depth and _pad_nan helpers

Test plan

• 23 tests for fused_overlap (correctness vs sequential, 1/2/3 stages, non-square depth, graph size reduction, numpy fallback, validation, accessor)
• 16 tests for multi_overlap (correctness vs sequential stack, shape, band dim, coords/attrs, dtype, graph size reduction, boundary modes, numpy fallback, validation, accessor)
• Accessor expected-methods tests updated
• All 74 tests pass, no regressions