Added alphabet check before decoding #7

Resolves #6

Linear time complexity solution, guaranteed to work.

Alternatively you could check the expected length of the output matches but then you will have to deal with edge cases (such as the test case given).

I'm open to updating this, but it's so old an relied on so heavily by so many projects and devices (including the AppleTV SDK) that I think I'll have to do a major version bump even for a bugfix.

You could actually make this quite a bit faster by checking character ranges.

// off the cuff, not correct code
if (!(x >= "a".toCharCode() && x <= "Z".toCharCode() /* && range of symbols*/)) {
   throw Error('barf');
}

Also, I want to stick with Vanilla JavaScript (ES5) so that, for those who update, they don't have to introduce a compiler or change their other build tools.

I changed it to a range check (good advice).

I think I've changed it to ES5, but I'm not super familiar with what is ES5 and what came after so I could have missed things.

We use a package called localflare to run Cloudflare Workers for testing which depends on this package. I discovered the different behavior between atob and buffer when looking at b64url in our code.

For us making the change is beneficial because the closer we can get our environment to Cloudflare the better (and I'm really against Node's behavior on this) but I also understand being cautious about versioning. I'll just have to make a PR to localflare to update their version :)

(Wrangler is on our roadmap but it will be a little while before we get there)

Thanks.
I've scheduled review, update, merge, and publish this on Saturday.