Measure anisotropic MPO patch scaling to chi 900

[shinaoka]

Summary

• add a validated multiscale constructor for axis-aligned and randomly rotated anisotropic Gaussian QTT MPOs
• make large Gaussian-mixture TCI input generation deterministic and scalable with center-derived pivots and a spatial evaluator whose omitted positive tails are globally bounded by 1e-12
• apply the input pipeline TCI 1e-8 -> global relative-L2/SVD 1e-6 -> adaptive patch cap 128
• record input rank, parameters, sampled errors, patch count/rank/parameters, cache metadata, and preparation time through compressed input chi 900
• add generated EPYC input-scaling records and report

Results

N	compressed input chi	patches	max patch chi
512	89	1	84
4,096	256	6	124
20,480	458	30	121
100,000	900	120	121
125,000	900	120	123

The patch count is approximately quadratic in compressed input chi over the measured range. All recorded patches remain below cap 128, and sampled input relative errors remain below 1e-5.

Verification

• cargo test --release (47 passed)
• cargo clippy --locked --all-targets -- -D warnings
• cargo check --locked --all-targets
• cargo fmt --all -- --check
• python3 scripts/report_mpo_input_scaling.py result/linux-epyc-7713p
• python3 -m py_compile scripts/report_mpo_input_scaling.py
• README local-link validation
• git diff --check
• independent review: Correct-to-merge