fix: avoid scanning cwd during server startup

[mikemikimike]

Summary

Keep the shared SDContextParams::lora_model_dir default as . so sd-cli continues resolving prompt LoRA tags from its current working directory. sd-server clears that default before parsing command-line arguments, preventing a recursive scan of its working directory unless --lora-model-dir is explicitly supplied.

• Preserve CLI prompt LoRA loading from the current working directory.
• Keep server LoRA discovery disabled by default.
• Honor an explicit server --lora-model-dir value.

Issue: #2081

Fixes #2081

Verification

• cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DSD_WEBP=OFF -DSD_WEBM=OFF -DSD_SERVER_BUILD_FRONTEND=OFF -DGGML_NATIVE=OFF
• cmake --build build --config Release --parallel 2
• ctest --test-dir build --output-on-failure (no tests are registered in this repository)
• build/bin/sd-server --help
• build/bin/sd-cli --help
• clang-format --dry-run --Werror -style=file examples/common/common.h examples/server/main.cpp

[leejet]

SDContextParams is shared by sd-server and sd-cli. Changing lora_model_dir from "." to an empty string also causes CLI generation to skip extract_and_remove_lora() when --lora-model-dir is omitted.

For example, sd-cli -m model.safetensors -p "a cat <lora:foo:1>" previously loaded foo.safetensors from the working directory. With this change, the LoRA is silently ignored and the tag remains in the prompt. This also contradicts the documented default in docs/lora.md.

Could we preserve "." in the shared parameters and clear it only during server initialization, before parsing command-line arguments? That would avoid the default server scan without breaking existing CLI usage.

[leejet]

Thank you for your contribution.