fix: decode split chunks after readable setEncoding

[marko1olo]

Fixes #5002.

Summary

• route BodyReadable.setEncoding() through Readable.prototype.setEncoding() so Node installs its StringDecoder
• preserve the previous invalid-encoding no-op guard
• add a regression for a three-byte UTF-8 character split across response body chunks

Tests

• NODE_PATH=C:\hades\oss\undici\node_modules node --test test\readable.js
• NODE_PATH=C:\hades\oss\undici\node_modules C:\hades\oss\undici\node_modules\.bin\eslint.cmd --no-cache lib\api\readable.js test\readable.js
• git diff --check

[marko1olo]

Follow-up on the red CI after the amended 04f3c290 push:

• The readable/setEncoding regressions pass locally:
  • NODE_PATH=C:\hades\oss\undici\node_modules node --test test\readable.js -> 14 passed
  • NODE_PATH=C:\hades\oss\undici\node_modules node --test test\client-request.js --test-name-pattern "request multibyte" -> 47 passed
  • scoped eslint and git diff --check passed
• The current failing CI cluster is test/interceptors/cache.js (vary headers are present in revalidation request, max-stale, stale-if-error). I reproduced the same 49 passed / 3 failed result on a detached origin/main worktree at 6ea54ef8, without this PR patch.

So the current cache-interceptor failures look like a baseline/main failure rather than a regression from this readable change.