#2786 docs: the 0.1.44.1 changes no page told: a crashed engine start names itself, and a setup run again keeps the installed size

说明

## Summary

Docs only: the 0.1.44.1 changes that no page told - the server start error now names an engine crash and the CPU features it lacks (#2324), and a setup run again keeps the installed model size and the user own env entries. Every statement verified against the 0.1.44.1 code: serve/server.py (d981a51, 9aa674e), setup.py (de9115f), src/kernels/cpu/nvfp4_avx2.cpp and nvfp4_avx512.cpp (0b6fc1c).

## What changed

- **docs/TROUBLESHOOTING.md**: new entry under While installing or starting: the engine dies the moment it starts, before writing anything to its log. What the start error says since 0.1.44.1 (the signal / NTSTATUS wording, the missing AVX / AVX2 / FMA / F16C note on Linux, the launcher-own-lines note), why a crash is no longer retried as a GPU-runtime start, and what to report with the CPU model.
- **docs/DETAILS.md**: the matching row in the troubleshooting table, next to the engine stopped unexpectedly row.
- **docs/INSTALL.md**: Older CPUs - on 0.1.44 the ready-made engine could die with an illegal instruction before refusing an old CPU (the AVX/FMA tables ran at program start; fixed in 0.1.44.1); Where things are stored - a setup re-run now keeps `STRATA_MMQ_KQ`, `STRATA_HC_Q8` and `STRATA_ROUTE_TAIL_SKIP` outside Unsloth configs; Setup questions - the re-run pre-selects the installed size (marked "(installed)") and `--yes` takes it.
- **docs/AI_SETUP.md**: the note that `--yes` on a setup run again over an installed model keeps that model size.

## Extra Notes

Left alone, already accurate on this base: the Responses/Codex reasoning_effort row in DETAILS.md (updated by the 0.1.44.1 commit itself). The Metal thread-limit fixes (#2323), the CUDA 12.1 build fix and the pool_tasks_test skip have no user-visible doc claims. No code, no README, no translations touched.