GEMM tuners: an interrupted run loses everything (results are only written at the end)

Tuners write -o / -o2 only after every shape completes. A long run that is interrupted — machine reclaimed, container stopped, OOM, timeout — loses all results, including the shapes that had already finished.

On a shared machine this is the dominant cost of tuning: we lost multi-hour --libtype all runs repeatedly and ended up wrapping the tuner in an external chunking driver (6 shapes per invocation) purely to bound the loss.

Proposal: append each shape's winner (and its -o2 candidates) as it completes and flush; on startup, skip shape keys already present in the output file. That is consistent with how --compare/--update_improved already reads the tuned file, gives --resume for free, and makes CI's op_tune.sh restartable.

Found while tuning GLM-5.2 on gfx950 (see #5262 for a related scalability fix from the same campaign).

Start at the GEMM tuner code that writes the -o and -o2 output files, then compare it with the existing --compare/--update_improved reading path. Trace how shape keys are loaded and how CI invokes the tuner through op_tune.sh; done means completed shapes are appended and flushed, reruns skip keys already present, and interrupted runs retain earlier results.