Let's say everyone did as you propose and use usernames for salts. I could pre-compute dictionaries for usernames like root, admin, etc. and likely gain full access to a wide range of systems.
Our aim is to build a system where brute-forcing is always the cheapest option. Once that's the case, we're in a position to control "how expensive" it is for someone to invert even one derived key.
Imagine if UNIX systems did as you propose. The rainbow table for the salt "root" would let me do plenty of damage.
I am not sufficiently familiar with KDFs but I know that rainbow tables exist to crack straight up md5 hashed passwords for common usernames and passwords where the username has been used as the salt.
Maybe harder to do with a KDF because it's prohibitive to compute that many keys in the first place?
What we're trying to protect against is someone in possession of the data (such as Mega themselves, or anyone working for them, or hacking them, or seizing their assets, or just someone over the internet making an offline attempt against an account after discovering the auth hash) from making a brute force attempt to learn the plaintext content of the accounts.
So to help me crack accounts, I could make a database of outputs from the KDF, with every password combination pre-computed. "Start with "aaaaaa" as a password, and go all the way through "ZZZZZZ" out to some number of characters. The database would have 32 output bytes per entry. The first 32 bytes would be the output of "aaaaaa" through the kdf, then "aaaaab", and so on.
If my math is right, going out to 6 characters using letters numbers and punctuation, the database (probably just a file really) would be about 5 TB in length, and based on my single slow CPU's AES speed, would take about 99 CPU days to compute. Of course if you have 100 CPUs (or EC2, or a botnet) you could have it done in time for dinner tomorrow. Or just wait because someone will likely precompute one you can bittorrent in the next few days.
So, if I wanted to make a brute force attack against Mega's database of users, I would use this pre-computed database, reading it sequentially from disk, taking every 32 byte section and trying it as a key to decrypt the next layer of keys for a user. My CPU could likely check answers faster than the data would stream in from the disk.
So I would be able to brute force any combination out to 6 password characters in a few minutes using only the spare hardware sitting around my office and unoptimized code. Imagine what NSA could do.
So, finally getting back to answering your question. If the salt were used, then I couldn't pre-compute the KDF output. I would have to do it separately for each user. So then my attempts would (again on my weak CPU) proceed at only about 520 attempts per second. That's still about 500 times to fast. So, the design needs both a slower KDF, and salt. :)
Trivia: Want to time your own CPU and how many AES cycles it can do?
Hopefully that makes sense!