fix: handle filesystem errors during lora and upscaler cache scan (#1920)

[Yi-111-a]

Resolves #1920.

Problem

In refresh_lora_cache and refresh_upscaler_cache (examples/server/runtime.cpp), directory traversal checked entries using unguarded entry.is_regular_file(), entry.file_size(), and unhandled directory iteration.
When the directory contains restricted entries, broken symlinks, special mounts (e.g. /proc/1/map_files inside Docker containers or systemd overlays), entry.is_regular_file() throws an uncaught std::filesystem::filesystem_error (Operation not permitted / Permission denied).
This escaped to the HTTP server handler, causing /sdcpp/v1/capabilities and /sdcpp/v1/img_gen to fail with 500 Internal Server Error and breaking the WebUI.
Additionally, the runtime stage of the Dockerfiles did not set WORKDIR, defaulting to / where --lora-model-dir . scanned the entire Linux root filesystem.

Fix

1. In examples/server/runtime.cpp:
   • Use the std::error_code overloads for is_regular_file, file_size, last_write_time, and absolute.
   • Use fs::directory_options::skip_permission_denied for upscalers directory iteration as well.
   • Wrap scanning in try ... catch (const std::exception& e) with warning logs so an inaccessible entry or directory never crashes the request handler or returns a 500.
2. In docker/Dockerfile*:
   • Add WORKDIR /sd.cpp in the runtime stage of all Dockerfiles so the working directory defaults to /sd.cpp rather than container root /.

Fixes #1920