Accumulate ParallelLinearQubitOperator matvecs in-place (#1410)

[rosspeili]

Implements the suggestion from #1410: avoid functools.reduce(numpy.add, …) when combining partial matvecs, which was allocating a new full-sized array for every partial sum.

Used the approach from GCA and went a bit further:

• Shared in-place += accumulation for both the single-process and multiprocess paths (same cost when combining worker results).
• Only set forkserver when processes > 1, so the single-process path works on Windows.

There is one nuance with processes=1, where operator grouping usually yields a single group so the biggest + is on the multiprocess reduce. Happy to trim or adjust if its better to keep the change scoped only to the single-process path.

Tests passed, and will wait for CI to see if I missed anything.

Fixes #1410

[gemini-code-assist]

Code Review

This pull request optimizes memory usage and improves platform compatibility for ParallelLinearQubitOperator. It replaces functools.reduce(numpy.add, ...) with a new helper function _accumulate_vectors that performs in-place vector accumulation to avoid intermediate allocations. Additionally, it ensures that the 'forkserver' multiprocessing start method is only set when spawning multiple processes, allowing the single-process path to run on platforms like Windows. Unit tests have been added to cover the new helper function and the single-process execution path. There are no review comments, so I have no feedback to provide.