Any language can be recognized by a Turing machine with an infinite number of states. Let the state graph be an infinite trie; scan over the input from left to right once, and you end up at a unique state for each possible input string.
By contrast, if you are familiar with different cardinalities of infinite set, it is easy to see that there must be some language not accepted by any Turing machine with a finite number of states. There are uncountably many languages and only countably many finite Turing machines. This argument works for any method of describing languages using descriptions of finite size: no such method can give a description for every language.
The Church-Turing thesis, according to Wikipedia, is the statement that
a function on the natural numbers can be calculated by an effective method if and only if it is computable by a Turing machine.
This does not contradict that a "Turing machine with infinite states" could recognize any language, because the phrase "Turing machine" in unqualified form always refers to a machine with a finite number of states.
The very fact that extending the Turing machine model to permit an infinite state set would allow it to recognize any language means that such a model is mathematically almost useless. Saying that a language is recognized (in linear time!) by a "Turing machine with infinite states" adds no information, because that is true of every language.