fix: the agent can read the global and built-in skills it lists

[mekjr1]

Fixes #37.

The skill catalog in the system prompt tells the model to read each skill's SKILL.md with read_file and gives its absolute path. With Restrict to Workspace on, the default, read_file reads only the workspace, the media folder and tools.allow_read_paths. It answered "path escapes workspace" for a skill in $COMPA_HOME/skills or in the built-in folder ($COMPA_BUILTIN_SKILLS, or skills beside the executable).

buildAllowReadPatterns now adds the global and built-in skill folders, as it adds the media folder. It uses the same ^<folder>(?:<separator>|$) pattern, which the file tools also check against the link-resolved path. The context builder and the patterns get the global folder from one helper, globalSkillsDir, and the built-in one from skills.BuiltinDir. So they name the folders the loader lists.

Every tool that takes the read paths uses these patterns: read_file, list_dir, exec, send_file, load_image and the message tool's local media. So a skill's other files, and the scripts its instructions run, can be reached too. The writing tools use tools.allow_write_paths and are unchanged. Embedders no longer need COMPA_TOOLS_ALLOW_READ_PATHS for their built-in skills.

docs/use.md now names the skill folders among the paths file tools may read and commands may name. It also names the attachment folder for commands, which was already allowed. CHANGELOG line.

Test: TestNewAgentInstance_ReadsTheSkillsTheCatalogLists puts one skill in each folder and reads every skill the loader lists with the agent's read_file. On main it fails with "path escapes workspace".

[mekjr1]

Superseded by #63, which includes this change or replaces it.