Android 4.0 doesn't like the `String.fromCharCode.apply` hack.

Avoids a potential race condition when the load event already fired
before a handler was attached. This was reported happening in Safari.

Fixes #353

Before, when a Request or Response was initialized with an ArrayBuffer
or TypedArray, that data was only accessible through `arrayBuffer()`
method, and if it was initialized with DataView, it wasn't accessible at
all.

Now, regardless of the original type of body, the accessors `blob()`,
`arrayBuffer()`, and `text()` always work.

This is in accordance with Section 6.4 of the spec:

    dictionary ResponseInit {
      unsigned short status = 200;
      ByteString statusText = "OK";
      HeadersInit headers;
    };

Fixes #376

Support ArrayBufferView types as POST body

Fixes #350.

This will get all the ArrayBuffer & Blob tests passing.

PhantomJS 1.9.8 seems to support ArrayBuffer & DataView, but not Blob.

DataView is one of the possible views for ArrayBuffer, which the fetch
spec says should be supported. However, IE 10-11 don't handle it well
when passed unmodified to XMLHttpRequest. As a workaround, wrap every
DataView in a Blob.

It's interesting to note that IE 10's string representation of a
DataView instance is `[object Object]` rather than `[object DataView]`.
Therefore, the only way to detect it is via `isPrototypeOf`.

Section 6.2 of the spec says that body can be BufferSource.
BufferSource is defined in https://heycam.github.io/webidl/#BufferSource:

> The BufferSource typedef is used to represent objects that are either
> themselves an ArrayBuffer or which provide a view on to an ArrayBuffer.

Furthermore, passing an ArrayBuffer directly to XMLHttpRequest is
considered deprecated and Safari throws a warning about it. This makes
it even more important to support ArrayBufferView in fetch.

The fallback for `ArrayBuffer.isView` is provided for IE 10, which
doesn't implement the method.

Spec compatibility for Request/Response constructors and cloning

Body of Request must be cloned using a value to avoid consumption of
it. Once a new body is initialized, re-initialize the old value.

Fixes #308, closes #335

Now splits by `\r\n` because that's the header line delimiter.
No tests because I can't figure out how to return a manually formatted
headers payload from Node.js HTTP server.

Fixes #422

Fixes #411

Section 6.4 of the spec:

> If init’s `headers` member is present, run these substeps:
> ... Fill r’s Headers object with init’s `headers` member.

Fixes #416

Section 6.5 of the spec:

> Let `request` be the associated request of the result of invoking
> the initial value of Request as constructor with `input` and `init`
> as arguments.

The Headers, Request, and Response suites should mainly be exercising
those classes and not invoking the `fetch` method. Most of those tests
that do have now been moved to the "fetch method" suite.

Section 6.3 of the spec:

> If `input` is a string, then run these substeps: ...
> Otherwise (`input` is a Request object), run these substeps: ...

It's worth noting that Chrome and Firefox don't implement this right
now, and will cast any non-Request object passed as 1st argument (even
null/undefined) to a string.

Improve Readme

Hopefully this will help reduce the number of invalid issues opened.

No longer recommend es6-promise due to its size and the fact it doesn't
conform to Promises/A+. Fixes #415

Fixes #393

Fix grammar

The adjective similar is modifying behave instead of a noun or pronoun. Use an adverb to modify a verb, adjective, or other adverb.

Fixed typo in README.md

Microsft -> Microsoft

It's been broken for a while, causing everyone's PR to show up as a
failed build. This isn't nice to our contributors, since it's not their
fault.