TensoRT doesn't cache plans during synchronous training loop even if use USE_CACHE_TENSORRT_PLAN

[KvanTTT]

Every time I run selfplay or gatekeeper it spends a lot of time during creating plan (up to 5 minutes):

2025-10-02 10:34:02+0200: TensorRT backend thread 0: Initializing (may take a long time)
2025-10-02 10:34:04+0200: Creating new plan cache
2025-10-02 10:38:22+0200: Saved new plan cache to \katago_contribute\katagodots-test\scripts\dated\20251002-032522\bin/KataGoData/trtcache/trt-101100_gpu-771cbe50_net-katagodots-test-dotsgame-s3612416-d363241_7_max12x12_batch128_fp16

And the trtcache directory is empty after iteration.

Also, I see the plan's name is bound to a certain model like s3612416-d363241. Is it possible to make it shared over generated models?