[XLA:CPU] Optimize Gather with non-contiguous slice sizes

[copybara-service]

[XLA:CPU] Optimize Gather with non-contiguous slice sizes

Optimizes gather with non-contiguous slice sizes using two complementary approaches:

1. Gather-Transpose prepacking in GatherSimplifier: Allows constant folding to hoist the transpose completely out of loops when weights are static.
2. Contiguous priority heuristics in LayoutAssignment: Defends against cache thrashing in dynamic buffers by guiding the layout assignment heuristic to assign a column-major memory stride for corresponding column-major slice gathers.