Fix canonicalize_url IDNA-encoding the port and userinfo of a non-ASCII host

[patchwright]

Fixes #222.

Problem

canonicalize_url runs IDNA encoding on the entire netloc, so when a non-ASCII host has a port (or userinfo), those parts get folded into the last IDNA label and mangled:

>>> canonicalize_url('https://тест.тест:33')
'https://xn--e1aybc.xn--:33-qdd4dec/'   # the :33 is corrupted

Expected:

'https://xn--e1aybc.xn--e1aybc:33/'

safe_url_string already handles host and port separately; only the canonicalize_url path (via _safe_ParseResult) was affected.

Fix

Add _idna_netloc, which applies IDNA to the host substring only and leaves any userinfo@ prefix and :port suffix verbatim. IPv6 literals are ASCII (IDNA is a no-op) so the bracketed host is returned untouched. Because IDNA is a no-op on ASCII, no existing ASCII-netloc behaviour changes. The existing UnicodeError fallback (too-long / missing labels) is preserved.

Tests

Added test_canonicalize_idns_with_port covering host+port, userinfo+port (userinfo stays case-sensitive), and the no-port case. Proven to fail on unpatched source (xn--33-mlc2cdc) and pass with the fix. Full suite: 359 passed, 0 failures; mypy + ruff clean.

Note: the trailing-multiple-dots addendum in #222 is left out of scope, as the reporter noted it may be standards-correct behaviour.