[leburgel]

Closes #596. Stacked on #581.

• fusiontrees(W) is a FusionTreePairs, an AbstractIndices that stores the splitting and fusion trees of every coupled sector instead of every pair. The position of a pair follows from those of its trees, and is its (integer) token.
• subblockstructure(W) is a FusionTreePairDict sharing those tokens.
• fusionblocks groups the trees of each side by their uncoupled sectors, combines the groups in the order in which they first occur (the order of Construct fusion tree blocks from the shorter side, and group the cached trees in fusionblocks #581), and fills every block once at its final size.
• degeneracystructure loops over the trees of both sides.

The transformers are identical to those of #581, block order included, on all transformers of one CTMRG iteration (554 for fℤ₂ × U(1) × SU(3), 550 for fℤ₂ × U(1) × SU(2) × U(1)).

Building those transformers, on top of main + #580 + #581 + #592 (D = 7, shared machine at load 16–29, ranges over 3 interleaved rounds):

	before	after
fℤ₂ × U(1) × SU(2) × U(1), serial preparation	15.6–17.3 s	9.9–11.1 s
fℤ₂ × U(1) × SU(2) × U(1), 1 thread	27.0–29.5 s	20.5–22.9 s
fℤ₂ × U(1) × SU(2) × U(1), 8 threads	18.9–20.0 s	14.1–17.9 s
fℤ₂ × U(1) × SU(3), serial preparation	4.4–4.7 s	3.4–3.7 s
fℤ₂ × U(1) × SU(3), 1 thread	10.6–10.7 s	9.5–10.3 s
fℤ₂ × U(1) × SU(3), 8 threads	4.9–5.5 s	4.2–4.6 s

Keying the groups by the integer positions of the uncoupled sectors instead makes fusionblocks about 1 s faster for SU(2), but the totals are within the noise, so this keeps the simpler version.

🤖 Generated with Claude Code