Periodic checkpoint output and lock-free hot-loop diagnostics

[krystophny]

Motivation

The tracer writes its result files (times_lost.dat, confined_fraction.dat, class_parts.dat) only in the final write_output. A run killed before completion, such as a batch-scheduler timeout on a long slowing-down run, therefore loses everything and offers no progress signal while it runs.

Separately, the symplectic integrators print and dump to fort.660x on every Newton non-convergence and every r<0 event. These sit in the inner integration loop, so a difficult case floods stdout and serializes the OpenMP threads on the implicit I/O lock.

Change

• diag_counters: per-thread, cache-line-padded event counters. The integrators call count_event(...) instead of printing; totals are a reduction over the thread columns, taken outside the hot path. No atomic and no critical on the increment.
• progress_monitor: once per finished particle a thread reads the wall clock, which is lock-free. Every checkpoint_interval seconds one thread enters a rarely taken critical, writes the partial results and a one-line status (percent done, ETA, aggregated event counts), and advances the deadline. A killed run keeps its last flushed output. The module holds a dump callback, so it depends on no tracing code.
• simple_main: write_output splits into a reusable write_results shared by the periodic checkpoint and the final write. The confined-fraction normalization moves there unchanged.
• New namelist key checkpoint_interval (seconds, default 10; set <= 0 to disable).

Verification

Build: cmake --build build (Release, gfortran 15), library and simple.x link clean.

Periodic output and recovery. A deterministic volume fast-classify with checkpoint_interval=1:

• status line every ~1 s: [progress] 10.00% 200/2000 t=19.1s eta=171.8s newton1_maxit=10 r_negative=10 (aggregated counts, no per-iteration spam);
• confined_fraction.dat, times_lost.dat, progress.dat present mid-run;
• after kill -9, the files remain; times_lost.dat holds the finished particles with -1 sentinels for the rest.

Final output unchanged. Same deterministic volume run with the monitor off (checkpoint_interval=0) and on (=10, run completes under one interval) produces a byte-identical class_parts.dat (diff clean). Confined fractions still divide by ntestpart; the per-particle files differ from the old code only by open(newunit=...) in place of a hard-coded unit.

The full golden_record suite runs in CI on this PR.