Does a practical collision attack on a cryptographic hash function also mean we can (or should) consider it to fail “indistinguishable from random data”? Or does a collision attack have no influence on that?
For a random oracle with output length $n$, it takes $2^{n/2}$ time to find a collision. So if you have a hash function with output length $n$ as well but for which there's an attack that finds a collision in less than $2^{n/2}$ time, then the hash function doesn't behave like a random oracle, period. The function has some property that a random oracle does not have.
Note that even before we ever found a collision, we already knew that Merkle-Damgård hash functions like MD5, SHA-1 and SHA-2 don't behave like random oracles, because they're vulnerable to length extension attacks: for any choice of message $m$, an adversary who knows the result of $\mathrm{hash}(m)$ and the message length $|m|$ can use this knowledge to "shortcut" the calculation of $\mathrm{hash}(m || \mathrm{pad}(|m|) || s)$ for any choice of $s$. A random oracle doesn't have this property—knowing the hash for any message is of no help for predicting the hash of any extension of it.
This doesn't mean that it's impossible to find some construction that uses the hash function in a carefully restricted way such that its output is indistinguishable from random (and indeed, HMAC-SHA1 is claimed to be a pseudorandom function—a function such that if you pick a secret key at random, the adversary can't tell it apart from a random function). It means that from the point of view of an adversary that has control over the input of the hash function, they can very easily tell that it doesn't behave like a random function would.
(Note that I have been careful to talk about random oracles and random functions instead of just the outputs of a function; the very stringent question we ask of a hash function is how it behaves from the point of view of an adversary who can choose what inputs to feed to it. Again, this doesn't rule out the possibility that in a weaker attack model—one where the adversary has fewer powers—the adversary might not succeed.)