Previously LoadIdentityFiles called ssh.ParsePrivateKey on every file
listed as IdentityFile in ssh_config and returned the parse error to the
caller on the first failure. That aborted the whole connection attempt
with a 'ssh: no key found' error bubbling up from inside the SSH dial.

The failure surfaced even when the SSH agent (e.g. yubikey-agent,
Secretive, 1Password, gpg-agent) was perfectly capable of covering the
requested identity. The most common trigger is a hardware-token setup
where the IdentityFile path points at a public-key-only file on disk
while the private half lives on the token and is accessed exclusively
through the agent. OpenSSH tolerates this and falls back to the agent;
pretty did not.

LoadIdentityFiles now:

  - Silently skips IdentityFile entries whose files do not exist
    (errors.Is(err, os.ErrNotExist)).
  - Silently skips entries that contain only an authorized_keys-style
    public key (detected via ssh.ParseAuthorizedKey), so hardware-token
    identities are delegated to the SSH agent.
  - Silently skips passphrase-protected keys
    (ssh.PassphraseMissingError), since this process has no way to
    prompt for a passphrase.
  - Still returns a wrapped, path-qualified error for unreadable files
    and for truly malformed private-key content so real
    misconfigurations are not hidden.

Tests:

  - Replace TestLoadIdentityFilesFailFast (which asserted the old
    behaviour) with TestLoadIdentityFilesSkipsMissing.
  - Add TestLoadIdentityFilesSkipsPublicKey for the hardware-token case.
  - Add TestLoadIdentityFilesMixedSkipsUnusable to verify that a usable
    private key alongside unusable entries still produces an auth method.
  - Add TestLoadIdentityFilesSkipsPassphraseProtected.
  - Add TestLoadIdentityFilesReturnsErrorForMalformedPrivateKey to pin
    the one case that must still surface an error.
  - Introduce a generateTestPublicKeyLine helper that synthesizes a
    fresh ed25519 authorized_keys line at runtime, so no static key
    material is checked into the tree.

Add focused tests for previously-uncovered branches in config.go:

- Error propagation from loadConfig for non-ENOENT paths (ENOTDIR via
  intermediate regular file) and from both user and system config loads
  in LoadSSHConfig.
- Resolver error propagation through getValue and getAllValues when the
  underlying library rejects an empty HostArg.
- ResolveHost paths for HostSpec.Alias precedence, invalid Port values
  surfacing as a wrapped error, and fallback to currentUser() when no
  explicit fallback is provided.
- LoadIdentityFiles wrapping non-ErrNotExist read errors with the
  offending path.
- ParseProxyJump dropping empty/whitespace-only components between
  commas.

Statement coverage for internal/sshConn/config.go: 84.61% -> 93.75%.
Remaining gaps are platform-specific (Windows shell branch in
shellMatchExec), environment-dependent (currentUser env fallbacks when
user.Current() fails), or structurally unreachable under normal control
flow.