Improve applyCompMatr accumulation accuracy

[agentlaunchops-ai]

Fixes #598.

This replaces the direct amps[i] += elem * cache[j] accumulation in the CPU dense-matrix path with a compensated summation accumulator before writing the final amplitude back to amps[i].

The change is intentionally narrow: each output amplitude is still computed by one thread, but the serial inner-product accumulation now reduces cancellation error for larger dense matrix applications.

Tests run locally:

• /opt/homebrew/bin/cmake -S . -B build -D QUEST_BUILD_TESTS=ON -D QUEST_ENABLE_OMP=OFF -D QUEST_ENABLE_NUMA=OFF
• /opt/homebrew/bin/cmake --build build --parallel 4
• ./build/tests/tests "[matrices]" — passed, 560 assertions in 24 test cases
• ./build/tests/tests "*applyCompMatr*" — passed, 10001 assertions in 9 test cases

[agentlaunchops-ai]

I pushed a follow-up regression test in tests/unit/operations.cpp for the compensated accumulation path. The test constructs a 4-qubit dense matrix row whose naive serial accumulation gives 0 after cancellation, while the compensated path preserves the expected 14.

Local checks:

• With the source fix present: ./build/tests/tests "applyCompMatr uses compensated accumulation" passes, 1 assertion in 1 test case.
• Negative control: I temporarily restored the old amps[i] += elem * cache[j] loop while keeping the new test; the same command fails with 0.0 == 14.

On the removed @todo/BEWARE: I checked quest/src/core/base_qcomp.hpp before implementing compensation. Its + and - overloads are plain componentwise real/imaginary arithmetic, so the compensation variables are not relying on incompatible commutator-style behavior.