perf(py): batch-path allocation and syscall nits

[dekobon]

Summary

Four allocation/syscall inefficiencies on the Python bindings' batch
path, found during the #1238 review. None is a correctness bug; all are
per-file costs on entry points now documented for large-tree use
(analyze_paths(root, skip_generated=False)).

Findings

1. Double stat per file — analyze_path
   (big-code-analysis-py/src/analysis.rs) runs path.is_dir()
   immediately before read_file_with_eol, whose first line is
   fs::metadata(path). Two stat(2) per file, every file, both entry
   points. The directory probe only matters when the read cannot
   produce Ok(Some), so it can move off the hot path — but the
   EISDIR-masking comment in analyze_path documents subtle platform
   behaviour, so the restructure needs care rather than a drive-by.
2. FuncSpace.name allocated before the gates — the owned path
   String is built ahead of both the read gate and the generated
   filter and dropped unused for every skipped file. The non-UTF-8
   check must stay ahead of the read (error-ordering contract); the
   allocation need not.
3. attach_or_keep clones the whole metrics JSON per file per
   injector (big-code-analysis-py/src/batch.rs) to have a fallback,
   even when the injector is a no-op (untracked file). With
   vcs=True, vcs_per_function=True that is two full-JSON clones per
   file. An injector signature of PyResult<Option<String>>
   (None = unchanged) removes the clone on both the no-op and success
   paths.
4. VcsRepoCache::resolve_root allocates two PathBufs per file on
   a cache hit — one to build the lookup key, one to clone the value.
   Arc<Path> values (refcount bump on hit) or a borrowed-key lookup
   removes both.

Non-findings (checked during the same review)

• py.None() per fix(py): batch docs promise 1:1 results under skip_generated=False, but tiny/binary files still drop slots #1238 placeholder is a refcount bump on an immortal
  singleton — nothing to hoist.
• No VCS machinery is touched for a declined file; the skip path costs
  one metadata call, one open, and at most a 64-byte probe read.
• analyze_batch's paths.len() preallocation is exact under
  skip_generated=false since fix(py): batch docs promise 1:1 results under skip_generated=False, but tiny/binary files still drop slots #1238; analyze_paths' now accounts for
  missing seeds too.