client: distinguish malformed request from network/server errors in Exchange

[jamiesun]

Summary

client.Exchange() previously returned a bare error for every failure, so callers could not tell a malformed/un-encodable request apart from a network or server problem. That distinction matters for failover logic: a malformed request should not be retried against the next server, whereas a connection error should.

This adds a typed MalformedRequestError (with Unwrap) and wraps only the request-encoding failure in Exchange with it. Network errors continue to satisfy net.Error, and context errors remain context.DeadlineExceeded / context.Canceled, so callers can now branch cleanly:

resp, err := client.Exchange(ctx, req, addr)
var malformed *radius.MalformedRequestError
switch {
case errors.As(err, &malformed):
    // bad request, do not failover
case err != nil:
    // network/server error, try the next server
}

Changes

• errors.go: add MalformedRequestError (Error, Unwrap) with usage doc.
• client.go: wrap the encode failure path; document Exchange error types.
• client_test.go: add TestClient_Exchange_malformedRequest.

No behavior change for existing callers that only check err != nil.

Closes #104