Fix .args loading order for program mode detection

[vlasky]

Note: This PR targets release v0.9.3 (commit ff0c02e) to make it easy for users relying on the last stable release to apply this fix to their local builds. Upstream main has since restructured to use submodules and patches.

Summary

• Move cosmo_args() before determine_program() so embedded .args flags like --server --v2 are visible when determining program mode
• Rename lf::server::main() to lf::server::run() with args_already_loaded parameter to avoid double-loading .args

Problem

When a llamafile has an embedded .args file containing --server --v2, the determine_program() call happens BEFORE cosmo_args() loads the embedded args. These flags are not seen, so the program falls through to chatbot mode instead of launching llamafiler.

Issue with PR #788

PR #788 correctly identifies this bug and moves cosmo_args() before determine_program(). However, it introduces a double-loading issue: when the program dispatches to llamafiler mode, lf::server::main() in prog.cpp calls cosmo_args() again. This causes .args to be loaded twice, which is problematic for accumulator flags like --header that append values rather than overwrite them.

Solution

This PR:

1. Moves cosmo_args() call before determine_program() in llama.cpp/main/main.cpp
2. Renames lf::server::main() to lf::server::run() with an args_already_loaded parameter
3. The dispatcher passes true to skip the redundant .args load
4. Standalone llamafiler binary passes false to load its own args

This provides the same fix as PR #788 while avoiding the double-loading issue.

Test plan

• Build llamafile
• Embed a .args file containing --server --v2 in a llamafile using zipalign
• Run the llamafile and verify it launches llamafiler instead of chatbot
• Verify standalone llamafiler binary still works correctly

Enhances the .args timing fix from PR #788.