BQRRP MEX: bulk memcpy marshal, factor in output buffers, static libstdc++

[mmelnich]

Performance and buildability fixes for the BQRRP MEX (release-plan item C2). Two commits:

1. Statically link libstdc++ on GNU/Linux. MATLAB ships its own older libstdc++.so.6, which lacks the GLIBCXX symbols emitted by GCC 13+, so a MEX built with a modern toolchain fails to load (GLIBCXX_x.y.z not found). This is the same treatment the funnystrompp branch already carries; main needs it to be buildable with current GCC at all. GCC/Linux-only; the flag list is empty for other compilers (AppleClang rejects -static-libgcc, and macOS has no GLIBCXX problem).

2. Bulk memcpy marshal; factor directly inside the output buffers. The input marshal bound a non-const TypedArray, which triggers MATLAB's copy-on-write unshare (a full deep copy of A before a single element is read), then copied element by element into a value-initialized std::vector (a third full pass, zeroing memory that is immediately overwritten). It is now a single memcpy from a const TypedArray view. In implicit mode the input is marshaled straight into the out1 MATLAB buffer and BQRRP factors in place, writing tau and J directly into their output buffers, so no scratch allocations or output round-trips remain; explicit mode keeps one non-value-initialized scratch (R extraction and ungqr post-process the factored form) and writes J in place.

Measured at 20000x500, implicit mode, median of 7 calls, interleaved old/new runs: 948 ms to 855 ms end to end (the ~93 ms delta is the eliminated copy passes over the 80 MB input plus the output round-trip). Correctness: test_bqrrp passes all six size/type cases (double/single at three shapes).

This is the same copy_into insight already applied to the funnystrompp MEX on the add-funnystrompp-mex branch (where the marshal cost was measured to dominate at n = 3000); PR #1 will merge cleanly over it since the CMake block is byte-identical.