[io] Fix a bug in concatenateTypedArrays

[caisq]

Previously, it was incorrectly assumed that the buffer of a
TypedArray has the same byteLength as the TypedArray itself.
The two can actually be different if the TypedArray is created
from an offset and a length from an ArrayBuffer, e.g.,
const x = new Float32Array(myBuffer, 32, 4).
This actually happens during the desrialization of a tf.Model.

This PR fixes it by checking whether the byteLength and
buffer.byteLength of a TypedArray are equal, and if not, creating
a new TypedArray with the buffer that matches the actual byteLength.

BUG

---

This change is 

[nsthorat]

Reviewed 2 of 2 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @caisq, @nsthorat, and @dsmilkov)

---

src/io/io_utils.ts, line 135 at r1 (raw file):

        x.byteLength === x.buffer.byteLength ? x :
                                               new (x.constructor as any)(x));
    if (!(x as any instanceof Float32Array || x as any instanceof Int32Array ||

I think you can you get rid of the any's here by using the following a few lines above:

```xs.forEach((x: TypedArray) => {`

<!-- Sent from Reviewable.io -->