Test extracts_through_deep_ambient_and_created_path in archive-trait reaches file descriptor limits

Working on a typical Fedora 44 Workstation installation:

```
$ git clone https://github.com/astral-sh/tar-codec.git
$ cd tar-codec/crates/archive-trait
$ cargo test
[…]
test extracts_through_deep_ambient_and_created_path ... FAILED

failures:

---- extracts_through_deep_ambient_and_created_path stdout ----

thread 'extracts_through_deep_ambient_and_created_path' (1495809) panicked at crates/archive-trait/tests/extract.rs:188:10:
deep destination should be removed: Os { code: 24, kind: Uncategorized, message: "Too many open files" }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

failures:
extracts_through_deep_ambient_and_created_path

test result: FAILED. 8 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.08s

error: test failed, to rerun pass `--test extract`
```

Running this under `strace` confirms that directories named `segment` are created, and their file descriptors are held open, until the default file-descriptor limit of 1024 is reached.

This happens in

https://github.com/astral-sh/tar-codec/blob/8fac35899b7e14727d61dc6687a4963f70cc28e5/crates/archive-trait/tests/extract.rs#L146-L156

and/or

https://github.com/astral-sh/tar-codec/blob/8fac35899b7e14727d61dc6687a4963f70cc28e5/crates/archive-trait/tests/extract.rs#L165-L169

It wasn’t immediately clear to me how to rewrite these loops to ensure the file descriptors of intermediate directories are closed once they are no longer actively needed, rather than consuming file descriptors proportional to the number of loop iterations.

Start with crates/archive-trait/tests/extract.rs, especially the loops at lines 146-156 and 165-169, then run cargo test --test extract to reproduce extracts_through_deep_ambient_and_created_path. Trace the directory creation and cleanup behavior, and confirm the test passes without holding intermediate directory file descriptors until the limit is reached.

Written by the indexing model from the issue text.