fix: don't panic when a source file disappears during diagnostic rendering

[MatthewMckee4]

Summary

utils::source_file is called from the diagnostic rendering path whenever we
need to attribute a span back to a test file on disk. It used to read the file
with .expect(\"Failed to read source file\"), which races against test files
being moved or deleted between discovery and diagnostic rendering and tears down
the entire worker process:

pub(crate) fn source_file(path: &Utf8Path) -> SourceFile {
    SourceFileBuilder::new(
        path.as_str(),
        std::fs::read_to_string(path).expect(\"Failed to read source file\"),
    )
    .finish()
}

It now falls back to an empty-bodied SourceFile via unwrap_or_default() —
the filename is still shown in diagnostics, but spans that reference the missing
file render as empty instead of crashing the run. A unit test exercises the
non-existent-path branch so the function is guaranteed not to panic on a missing
file.

This PR originally also tried to defuse two .expect(\"builtin fixtures to not fail\") panics in package_runner::run_fixture. Those sites were removed
entirely by #645, which rewrote built-in fixtures to run as pure Python, so
there is no longer a "this fixture has no AST" case at the Rust layer. The
rebase onto post-#645 main dropped those changes.

Test Plan

• just test (818 tests passing)
• uvx prek run -a

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

✅ 1 untouched benchmark

---

_{Comparing fix/builtin-fixture-panics (dc023ed) with main (9d04e61)}