Semiring rework

[rmanhaeve]

Refactored the torch layer hierarchy. The main goals were making it easier to define custom semirings and cleaning up the deep method dispatch chains.

Layer hierarchy (before → after):

• CircuitLayer was a god-class with _scatter_forward, _scatter_backward, _safe_exp, _scatter_logsumexp_forward all baked in. Now it's a thin wrapper around torch.scatter_reduce that just takes a reduce: str.
• LogSumExpLayer subclasses CircuitLayer and overrides forward — instead of the old runtime if reduce == "logsumexp" branching.
• ProbabilisticCircuitLayer was also tangled into CircuitLayer. Now it's a separate branch under AbstractCircuitLayer, with ProbabilisticSumLayer and ProbabilisticLogSumLayer as clean subclasses (no more if log_space branching or reduce_fn callables).
• New GatherCircuitLayer for custom reductions: pads groups into a 2D tensor so users just provide a standard torch reduction (e.g. torch.nanmean) + a fill value. No scatter knowledge needed.

Custom semirings:

• Semirings are now (sum_reduce, prod_reduce, zero, one, negate) where reduces are strings (for scatter_reduce-backed ops) or callables (for custom gather-based ops). No more passing layer constructors.
• e.g. ("amin", "sum", float('inf'), 0.0, tropical_negate) for tropical semiring.

Utilities:

• Extracted scatter_logsumexp and gather_indices into utils.py.

Performance:

• No regressions, actually ~6-20% faster across the board. The speedup comes from flattening the method dispatch chains (the old code went 5 calls deep for logsumexp) and removing per-forward string comparisons/nan_to_num_ calls.

Tests & benchmarks:

• 25 tests, 100% coverage on klay.torch.
• Added tests/test_benchmarks.py using pytest-benchmark for per-commit regression tracking.