fix(docker): set musl pthread stack via PT_GNU_STACK (real fix for #60)

[ccomb]

Summary

• PR fix(docker): unblock 8 GB VM target — OpenBLAS musl pthread + RTS memory discipline #60 was a no-op. It diagnosed the cause correctly (musl's 128 KB default pthread stack vs. OpenBLAS DYNAMIC_ARCH Fortran kernels with large auto-arrays → SIGSEGV on the first BLAS3 call from MUMPS) but patched a block that the compiler dropped: the injected pthread_attr_setstacksize(&attr, 8 << 20) sits inside #ifdef NEED_STACKATTR, and blas_server.c itself does #undef NEED_STACKATTR unconditionally on Linux (line 103). Verified post-merge: strings /usr/local/bin/volca | grep pthread_attr_setstacksize returns 0 matches in the volca-with-frontend:latest built from fix(docker): unblock 8 GB VM target — OpenBLAS musl pthread + RTS memory discipline #60.
• Bug reproduces unchanged. User report: same exit 139 at the same point — [MATRIX] Pre-computing factorization for database 'agribalyse-3-2' — on a 16 GB host running the freshly-pruned-and-rebuilt image. OPENBLAS_NUM_THREADS=1 works around it (single-threaded path skips pthread_create), confirming the root cause is still musl's default thread stack.
• Real fix: set the default at link time via -Wl,-z,stack-size=8388608. musl reads PT_GNU_STACK->p_memsz at process start and uses it as __default_stacksize. Every pthread the binary creates with NULL attr — OpenBLAS workers, GHC RTS capabilities, anything else — starts with 8 MB. No OpenBLAS source patching, no entrypoint env var, no thread-count downgrade.
• Removes the no-op sed and grep guards from docker/Dockerfile.

Why not patch OpenBLAS properly instead

Two reasons to prefer the linker flag:

1. NEED_STACKATTR can't simply be #define'd: the second pthread_create(&blas_threads[i], &attr, …) site in goto_set_num_threads references an attr that isn't declared in that function's scope — defining the macro turns latent dead code into a compile error.
2. The linker flag covers every pthread in the binary, not just OpenBLAS's. RTS capabilities (+RTS -N) inherit musl's default too, so a Haskell-side BLAS call from a non-OpenBLAS worker would still have been at risk.

Test plan

• ./volca/docker-build.sh --with-frontend from volca-deploy/ produces a volca-with-frontend image.
• readelf -l dist/volca | grep -A1 GNU_STACK (or on the image's /usr/local/bin/volca) shows MemSiz = 0x800000 (8 MiB).
• docker run --rm -p 8080:8080 -v /home/ecobalyse/volca_data:/data volca-with-frontend --config /data/volca.toml server --password … no longer crashes at MUMPS factorization for agribalyse-3-2 without OPENBLAS_NUM_THREADS=1.
• Run a full impact computation on the loaded DB to confirm BLAS multi-threading is actually happening (no silent fallback).