Fix race in proxy-bcrypt test that crashes Windows CI

[X-Guardian]

ci (latest, windows-latest) failed on master after #1001 (run), with the tape process exiting mid-run during app1 build:

Error: ENOENT: no such file or directory, utime '...\Temp\npg-download\napi-v3\bcrypt_lib.node'
Emitted 'error' event on Extract instance

The error comes from test/proxy-bcrypt.test.js, which had already passed a few seconds earlier. The same job passed on the preceding PR run and on the next PR run, so this is a timing flake rather than a regression.

Cause

The download test resolved on the gunzip stream's end. At that point tar-fs has written bcrypt_lib.node but has not finished: it sets the file's times afterwards. The cleanup test then called rimraf(downloadDir) without awaiting it. On a slow runner the delete won, tar-fs's utime hit ENOENT, and with no error listener on the extract stream the error was unhandled and killed tape.

Changes

Resolve on tar-fs finish instead of gunzip end, and reject on errors from either stream, so the test only ends once extraction is complete and any extraction error fails the test rather than the process.

Await rimraf in cleanup, so the directory is gone before the next test file starts.

Create downloadDir in setup. It previously created a stray ./download directory in the working directory, which nothing used.