metro-file-map: Vendor walker and remove the dependency - #1905

## Summary
`metro-file-map` depends on `walker` for `FallbackWatcher`'s initial crawl and for nothing else - a single call site in `recReaddir`. It's ~120 lines of recursive `lstat`/`readdir`, unmaintained since 2021, untyped (we carry a `$FlowFixMe[untyped-import]` for it), and it pulls in `makeerror` and `tmpl` purely to construct one error class we never inspect.
This diff brings it in-tree at `packages/metro-file-map/src/third-party/walker.js`, following the `buck-worker-tool/src/third-party` convention - the Apache-2.0 notice retained as `LICENSE.APACHE2` alongside it, and a `Portions (c) Meta Platforms` header over Metro's own MIT notice. Behaviour is preserved: the same event set (`entry` followed by a type-specific event, plus `error` and `end`), the same `filterDir` semantics, the same pending-counter completion, and the same `lstat` dispatch order - which matters, because `isFile()` is tested last and only after the device types. It's modernised where Flow required it: an ES class composing an `EventEmitter` rather than `util.inherits`, `node:fs`/`node:path`, and a local `UnknownFileTypeError` subclass in place of `makeerror`.
`on()` is typed against a keyed listener map, so both the event name and the listener signature are checked at the call site. I checked that isn't vacuous - renaming `'dir'` to `'dirr'` in `recReaddir` gives `property 'dirr' (did you mean 'dir'?) is missing in EventListeners`.
Two consequences worth calling out:
 - Because `walker` required the unprefixed `'fs'`, `resolver-test.js` carried a `jest.mock('fs', ...)` aliasing it to the same in-memory fs as `node:fs`. The vendored copy uses `node:fs`, so that goes.
 - Apache-2.0 source is now in-tree in an MIT repo. It was already in the dependency graph and it ships either way, but this puts it in the tarball as our own source, so it's worth a look from that angle. `LICENSE.APACHE2` is copied to `build/third-party/` by `scripts/build.js`, which copies non-JS files under `src/` verbatim.
The file is under `third-party/` rather than `vendor/` deliberately. Both are in `.eslintignore`, but only `packages/**/third-party` is in `IGNORED_PATTERNS` in `generateTypeScriptDefinitions.js` - under `vendor/` the build fails at `generateTypeScriptDefinitions.js:164` reading `lintResult.output`, because ESLint returns no result at all for an ignored file. So `third-party/` is the name both configs already agree on, and this way needs no config change.
`walker` stays in `yarn.lock` under `jest-haste-map`, which is dev-only, so no published Metro package depends on it (or on `makeerror`/`tmpl`) any more.
Changelog: [Internal] Vendor the `walker` dependency into `metro-file-map`
## Test plan
```
yarn flow check
yarn typecheck-ts
yarn lint
yarn verify-api-snapshots
yarn build
yarn jest
```
All clean - full Jest run is 2662 passed, 1 skipped, across 147 suites.
`recReaddir` drives `FallbackWatcher`'s initial crawl, so `packages/metro-file-map/src/watchers/__tests__/integration-test.js` exercises the vendored code directly - 34 tests, including the `Fallback` matrix over new/changed/deleted files, symlinks (including to non-existent targets), pre-existing trees moved in and out of the watched root, and directory deletion. There's no dedicated unit test for `walker.js` itself.
One caveat: the first full run failed at `integration-test.js:92`, on the `NativeWatcher` directory-event assertion rather than `Fallback`. It passed in isolation and on a clean re-run of the whole suite, so it reads as a timing flake under parallel load, but I haven't bisected it against unmodified main to prove that.