Implement recursive decomposition (Schur-complement) log-det computation

[Balance-H]

Motivation

• Replace the previous flattened-atom approach with a nested, recursive evaluation of log|det Omega| that follows the decomposition formula (Schur-complement recursion) from graphical-model theory (Lauritzen) so the result matches the nested atoms/separators structure.
• Preserve existing eigvals and dad methods while enabling an exact decomp option that operates on the decomposition tree instead of simple per-atom aggregation.

Description

• Stop flattening the decomposition in main.py and cache the nested decomposition trees as graph_y.atoms_decomp and graph_y.seps_decomp, and print group counts using a _count_groups helper for diagnostics.
• Add log_det_method and decomp_log_det selection flags in FlexLayer and require atoms_decomp / seps_decomp when log_det_method == "decomp".
• Implement recursive Schur-complement log-determinant evaluation in layers/flex.py: construct the layer precision matrix Omega on the current device (respecting dist_weighted), split the nested decomposition into left/right atoms and separator, form the AA, BB, SS, AS, BS blocks, compute Schur complement blocks using Cholesky solves, and recursively evaluate -logdet(S) + logdet(left) + logdet(right) until atomic subgraphs are reached.
• Add robust index handling and fallbacks where a split or index mapping cannot be applied, and use torch.linalg.slogdet, torch.linalg.cholesky, and torch.cholesky_solve for stable factor/solve operations.

Testing

• No automated tests were executed for this change (environment lacked required dependencies such as torch/igraph).

---

Codex Task