Phase 1: Native-Fortran adapter + CUDA host setup #85

Part of epic #84.

Scope

Build native amica from the .f90 source (gfortran + OpenMP) on a Linux x86 + CUDA host, and add a _run_fortran native-Fortran backend to the Phase B dimsweep harness (benchmarks/benchmark_dimsweep.py).

Definition of Done

• Build script/docs for gfortran + OpenMP compile of amica15.f90/funmod2.f90 on Linux x86.
• _run_fortran adapter: writes .fdt (Fortran-order) + param file, invokes the binary, parses the binary's own per-iteration timestamps for startup-immune per-iteration timing (not wall-clock of the whole process).
• Registered as a backend in _BACKENDS with availability detection (present only when a native build exists).
• Structurally validated on Mac (Rosetta amica15mac) and against the reverted Phase D prototype; real timing deferred to the Linux host.
• Real EEG only (channel subsets of ds002718). NO MOCKS.