fix: report output write failures as one sqlcj diagnostic

[omer-cengel]

Summary

When sqlcj generate could not create an output directory or write a generated file, the failure escaped as an unhandled exception with a stack trace. It now ends the run with a single sqlcj: diagnostic naming the resolved path that could not be written, and exit status 1, matching how configuration and source failures are already reported.

Changes

• Report Cannot create output directory: <path> when a generated file's parent directory cannot be created, and Cannot write generated file: <path> when the file itself cannot be written; both use the resolved path and keep the underlying filesystem exception as the cause.
• Propagate these failures from the compiler as CompilationException, so the existing generate error handling prints one line and exits with status 1 without a stack trace.
• Add tests at the writer, compiler, and CLI subprocess levels that force failures portably by placing a regular file where an output directory is needed and a directory where a generated file goes.
• Document both diagnostics, the exit status, and that files already written by the failing run stay in place, in the configuration guide and README.

Scope and non-goals

• Analysis still completes for every configured entry before the first file is written, and writing remains non-transactional: no rollback, staging directory, or stale-output cleanup is added.
• The diagnostic names the failing path only; it does not add the filesystem's reason or a suggested correction.
• Generated file content, write order, other diagnostics and their exit codes, and --config path resolution are unchanged.