[Yue-Zhengyuan]

This PR allows simple update with multi-site MPO gates to select an allocator, following what @lkdvos did in #459 for nearest-neighbor gates.

There is another improvement on how an (OBC) gate applies on its two endpoints, similar to how we factor out reduced bond tensors before applying a nearest neighbor gate.

[lkdvos]

Overall looks good to me!

[Yue-Zhengyuan]

J1–J2 iPEPO simple-update benchmark

J1=1, J2=0.5, spin-1/2, unpurified iPEPO, Float64, 2×2 unit cell, and Δβ=0.01.
One full su_iter dt sweep contains 8 nearest-neighbor gates and 16 three-site MPO gates, applied on both physical legs.
Timings below are the minimum of 5 samples per implementation, with one complete SU dt step per sample.
main is aa1392e2; branch is 11b27b85 (su-3site-allocator).

Case	Time main	Time branch	Speedup	Memory main → branch
J1–J2 iPEPO / Trivial / D7	217.2 ms	192.9 ms	1.13×	553.2 → 190.0 MiB
J1–J2 iPEPO / Trivial / D11	804.4 ms	641.3 ms	1.25×	3278.9 → 1064.7 MiB
J1–J2 iPEPO / U1Irrep / D7	69.9 ms	54.6 ms	1.28×	186.6 → 82.7 MiB
J1–J2 iPEPO / U1Irrep / D11	320.5 ms	275.8 ms	1.16×	932.5 → 341.3 MiB
J1–J2 iPEPO / SU2Irrep / D7	47.8 ms	40.1 ms	1.19×	42.9 → 30.6 MiB
J1–J2 iPEPO / SU2Irrep / D11	69.6 ms	57.5 ms	1.21×	174.1 → 88.0 MiB

The measured speedup is modest, around 1.13–1.28×, corresponding to roughly 11–22% less elapsed time.

The saving is mainly on total allocated bytes, which fall by about 29–68%, depending on the symmetry and D. Although allocation counts rise slightly (about 2–5%).

Symmetry charges on virtual spaces

Symmetry	D	Charges
U1Irrep	7	Rep[U₁](0 => 3, 1 => 2, -1 => 2)
U1Irrep	11	Rep[U₁](0 => 5, 1 => 3, -1 => 3)
SU2Irrep	7	Rep[SU₂](0 => 1, 1 => 2)
SU2Irrep	11	Rep[SU₂](0 => 2, 1 => 3)

@lkdvos Looks good?

[lkdvos]

Definitely okay for me, do we need to check the buildkite failures or were they already there?

[Yue-Zhengyuan]

Should be irrelevant to this PR. The failure also occurs in @kshyatt's several other PRs; maybe due to recent release of TensorKit 0.17.3?