the u.Hash error can be safely ignored (panic) because multihash
only fails from the selection of hash function. If the fn + length
are valid, it won't error.

cc @whyrusleeping

working on dag modification structures, factored out the data format into an importer subpackage and added more ipns tests

This commit fixed the notoriously annoying "Malformed Public Key"
problem. The issue was that sometimes the byte representation of
the points (x,y in big.Int) generated would be one less byte than
expected. This is simply because (* big.Int) Write uses the least
amount of bytes needed for the int.

I instead changed the marshalling/unmarshalling to do exactly
what stdlib crypto/tls does: use `ellipctic.Marshal` which marshals
according to the ANSI X9.62 standard.

http://golang.org/pkg/crypto/elliptic/#Marshal
http://golang.org/src/pkg/crypto/tls/key_agreement.go#L214

```Go
// crypto/tls
ka.privateKey, x, y, err = elliptic.GenerateKey(ka.curve, config.rand())
ecdhePublic := elliptic.Marshal(ka.curve, x, y)

// ipfs/crypto
priv, x, y, err := elliptic.GenerateKey(curve, rand.Reader)
pubKey := elliptic.Marshal(curve, x, y)
```

((Warning: we're using `rand.Reader` directly, which we shouldn't
do, as it can be seeded. We should use a configured source, as
crypto/tls. Flagged in #143))

This makes me think we should re-use a lot of their datastructures
and functions directly (e.g. ecdheKeyAgreement)

Fixed: #135

cc @bren2010 @whyrusleeping

this commit dedicated to @whyrusleeping