Optimize release builds and preserve profiling symbols

[mveril]

Summary

• Optimize release builds with thin LTO, one codegen unit, and symbol stripping.
• Keep the profiling profile optimized while retaining full debug information and symbols.
• Document all Cargo profiles and the stable-Rust command for targeting the build machine's CPU. There is no redundant native profile.

Before/after benchmarks

Both variants use the same source and default features. "Before" uses Cargo's default release settings (lto=false, codegen-units=16, strip="none") through cargo --config overrides; "after" uses this PR's release profile. Benchmarks ran on the same 12-logical-CPU host with fixed Rayon thread counts.

End-to-end RHF: samples/ethanol/6-31g/calculation.toml, local basis fixture, downloads disabled, JSON output. The 1-thread data uses Hyperfine with 2 warmups and 10 runs per binary. The 4-thread data uses 10 alternating before/after pairs to control for run-order effects.

Rayon threads	Before, mean ± SD	After, mean ± SD	Time reduction
1	2.254 ± 0.022 s	2.155 ± 0.008 s	4.4%
4	0.785 ± 0.034 s	0.754 ± 0.036 s	4.0%

The JSON results are identical, including the RHF total energy of -153.9984834558699 Hartree. The release binary shrinks from 22.56 MB to 13.44 MB (40.5%).

MP2 kernel: The existing benches/mp2_timings.rs benchmark runs synthetic RHF MP2 at 80 AO functions, 26 occupied orbitals, and a fixed 1 GB memory budget. Each invocation warms up and reports the median of three blocked-contraction timings; the table summarizes five alternating invocations per variant. The benchmark also checks blocked and dense energies agree.

Rayon threads	Before blocked MP2, mean ± SD	After blocked MP2, mean ± SD
1	466.7 ± 4.6 ms	467.0 ± 5.1 ms
4	460.2 ± 15.5 ms	454.0 ± 8.0 ms

The MP2 differences are small relative to their variation, so this benchmark does not establish a kernel speedup. Four threads matter substantially for the full calculation; the original 1-thread-only benchmark was insufficient. Fat LTO with one codegen unit was tested first and regressed the 4-thread HF workload by roughly 27%; thin LTO was selected based on the multithreaded result.

Verification

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• Baseline and optimized cargo build --release --locked builds
• Baseline and optimized cargo build --release --bench mp2_timings --features bench-support --locked builds
• Identical HF JSON outputs; the MP2 benchmark's dense/blocked energy assertions pass
• RUSTIQ_MP2_SIZES=80 RUSTIQ_MP2_MEMORY=1GB RAYON_NUM_THREADS=4 for the MP2 comparison

The full Cargo test suite was not run to completion; validation focused on execution time as requested.