Smith waterman fallback for length > 512

[maxRN]

When I try to find matches on a haystack that contains words with length > 512 (that is 513 or longer), these words get ignored in matches. I have created a reproduction test case in this PR: #19
I'm not super familiar so there is probably a better way to write the test.

Context for anyone curious (problem occurred in blink.nvim): I was using the texlab LSP for writing my thesis and in its LSP suggestions it sends the references of my bibliography along with the entire abstract. So the suggestions get very long.

[Saghen]

We should fallback to an algorithm that doesn't scale O(nm) when n * m > threshold. We could use memchr, which we already have an implementation for, but we'd only be able to support a limited number of typos (0, 1 or maybe 2 max), which should be fine. We would also need a fallback implementation for getting matched indices but that can come later