feat: implement GPU UCCSD(T)

[dajiaohuang]

Summary

• route PySCF UHF references through a new GPU UCCSD solver
• evaluate unrestricted CCSD amplitudes and energy in an antisymmetrized spin-orbital representation, then return PySCF-compatible (t1a, t1b) / (t2aa, t2ab, t2bb) blocks
• add the generalized spin-orbital (T) correction on the GPU
• enforce the existing gpulim / mem_ratio contract before uploading in-core transformed integrals
• make the still-unsupported GHF path fail explicitly instead of silently returning None

Implementation notes

The AO-to-MO transformation remains in PySCF on the host, matching the existing conventional CCSD path. Transformed spin-orbital blocks, iterative CCSD contractions, DIIS amplitudes, and (T) contractions run through CuPy. This first unrestricted implementation is deliberately in-core; it reports the required and configured GPU integral footprints before upload when the calculation is too large.

Validation

• cucc/test/test_uccsd.py: 2 tests passed on an NVIDIA RTX 5070 Ti with PySCF 2.5.0 and CuPy 14.2.0, using ByteQC's actual BufferPool
• LiH+ / STO-3G doublet: CCSD correlation energy and (T) agree with PySCF within 1.1e-18 and 3.2e-22; all five unrestricted amplitude blocks agree within 1.7e-17
• triplet H2O / 3-21G: CCSD and (T) agree within 3.5e-10 and 3.1e-11; frozen-core validation agrees within 2.8e-10 and 2.5e-11
• python -m py_compile cucc/__init__.py cucc/uccsd.py cucc/test/test_uccsd.py
• ruff check --ignore N999 cucc/__init__.py cucc/uccsd.py cucc/test/test_uccsd.py
• git diff --check

Closes #1