Optimize MP2 with blocked AO-to-MO contractions

[mveril]

The previous MP2 implementation expanded a dense AO-pair matrix and retained complete occupied/virtual MO tensors. Replace it for RHF and UHF with four successive contractions blocked over active occupied orbitals, accumulating each block's energy immediately.

Add [mp2] memory_limit = "auto" (default) or explicit sizes such as "512 MiB". Automatic selection uses half of available system memory, with correct KiB conversion and a 512 MiB fallback. Preserve automatic/fixed selection in the Rust API and round-trip explicit sizes without losing integer bytes. Text reports include the chosen block, workspace budget, dense-method upper bound, and resident inputs; JSON schema v1 is unchanged.

The budget covers owned matrix payloads, including coefficient copies and panels. It excludes existing HF data, allocator overhead and matrix-kernel scratch storage, so it is not an RSS limit. Tiny budgets fail with a located, typed error before transformation allocation. Adding the configuration field requires Rust struct literals to specify it or use ..Default::default().

Validation:

• cargo test --workspace --locked --offline, followed by the focused MP2 suite after final adjustments.
• cargo test --workspace --no-default-features --locked --offline.
• cargo clippy --workspace --all-targets --all-features --locked --offline -- -D warnings.
• cargo fmt --all -- --check, Ruff checks and formatting checks for the reference tooling.
• pytest tools/reference against the built binary: 17 passed, including 8 MP2 comparisons. New frozen-core H₂O/6-31G and OH/STO-3G cases cover small and large budgets.
• Independent four-index integral checks and dense energy comparisons cover frozen orbitals, unequal spin coefficients, rectangular MO matrices, block boundaries, 1/4 threads, budget boundaries and numerical errors.

Synthetic RHF release benchmark, four Rayon threads, median of three runs after warm-up; input/ERI construction and SCF are excluded:

AO count	Occupied	Dense time	Blocked time (auto)	Speedup	Dense payload upper bound	Blocked payload
12	4	0.650 ms	0.250 ms	2.60×	133.5 KiB	74.2 KiB
24	8	2.010 ms	1.838 ms	1.09×	2.0 MiB	1.1 MiB
48	16	56.017 ms	33.847 ms	1.66×	30.9 MiB	18.0 MiB

With a 4 MiB budget, the 48-AO case selects three occupied orbitals per block and uses 3.4 MiB of matrix payload, but takes about 182 ms. Small blocks trade speed for memory by repeating AO reads. These synthetic results are machine-dependent, not a speed guarantee for molecular calculations.

Run cargo bench --bench mp2_timings --features bench-support; RUSTIQ_MP2_MEMORY, RUSTIQ_MP2_SIZES, and RAYON_NUM_THREADS select the budget, dimensions and thread count.