A public salt will not make dictionary attacks harder when cracking a single password. As you've pointed out, the attacker has access to both the hashed password and the salt, so when running the dictionary attack, she can simply use the known salt when attempting to crack the password.
A public salt does two things: makes it more time-consuming to crack a large list of passwords, and makes it infeasible to use a rainbow table.
To understand the first one, imagine a single password file that contains hundreds of usernames and passwords. Without a salt, I could compute "md5(attempt[0])", and then scan through the file to see if that hash shows up anywhere. If salts are present, then I have to compute "md5(salt[a] . attempt[0])", compare against entry A, then "md5(salt[b] . attempt[0])", compare against entry B, etc. Now I have n times as much work to do, where n is the number of usernames and passwords contained in the file.
To understand the second one, you have to understand what a rainbow table is. A rainbow table is a large list of pre-computed hashes for commonly-used passwords. Imagine again the password file without salts. All I have to do is go through each line of the file, pull out the hashed password, and look it up in the rainbow table. I never have to compute a single hash. If the look-up is considerably faster than the hash function (which it probably is), this will considerably speed up cracking the file.
But if the password file is salted, then the rainbow table would have to contain "salt . password" pre-hashed. If the salt is sufficiently random, this is very unlikely. I'll probably have things like "hello" and "foobar" and "qwerty" in my list of commonly-used, pre-hashed passwords (the rainbow table), but I'm not going to have things like "jX95psDZhello" or "LPgB0sdgxfoobar" or "dZVUABJtqwerty" pre-computed. That would make the rainbow table prohibitively large.
So, the salt reduces the attacker back to one-computation-per-row-per-attempt, which, when coupled with a sufficiently long, sufficiently random password, is (generally speaking) uncrackable.
Salt is traditionally stored as a prefix to the hashed password. This already makes it known to any attacker with access to the password hash. Using the username as salt or not does not affect that knowledge and, therefore, it would have no effect on single-system security.
However, using the username or any other user-controlled value as salt would reduce cross-system security, as a user who had the same username and password on multiple systems which use the same password hashing algorithm would end up with the same password hash on each of those systems. I do not consider this a significant liability because I, as an attacker, would try passwords that a target account is known to have used on other systems first before attempting any other means of compromising the account. Identical hashes would only tell me in advance that the known password would work, they would not make the actual attack any easier. (Note, though, that a quick comparison of the account databases would provide a list of higher-priority targets, since it would tell me who is and who isn't reusing passwords.)
The greater danger from this idea is that usernames are commonly reused - just about any site you care to visit will have a user account named "Dave", for example, and "admin" or "root" are even more common - which would make construction of rainbow tables targeting users with those common names much easier and more effective.
Both of these flaws could be effectively addressed by adding a second salt value (either fixed and hidden or exposed like standard salt) to the password before hashing it, but, at that point, you may as well just be using standard entropic salt anyhow instead of working the username into it.
Edited to Add: A lot of people are talking about entropy and whether entropy in salt is important. It is, but not for the reason most of the comments on it seem to think.
The general thought seems to be that entropy is important so that the salt will be difficult for an attacker to guess. This is incorrect and, in fact, completely irrelevant. As has been pointed out a few times by various people, attacks which will be affected by salt can only be made by someone with the password database and someone with the password database can just look to see what each account's salt is. Whether it's guessable or not doesn't matter when you can trivially look it up.
The reason that entropy is important is to avoid clustering of salt values. If the salt is based on username and you know that most systems will have an account named either "root" or "admin", then you can make a rainbow table for those two salts and it will crack most systems. If, on the other hand, a random 16-bit salt is used and the random values have roughly even distribution, then you need a rainbow table for all 2^16 possible salts.
It's not about preventing the attacker from knowing what an individual account's salt is, it's about not giving them the big, fat target of a single salt that will be used on a substantial proportion of potential targets.
Best Answer
Let's do a little math: Credit card numbers are 16 digits long. The first seven digits are 'major industry' and issuer numbers, and the last digit is the luhn checksum. That leaves 8 digits 'free', for a total of 100,000,000 account numbers, multiplied by the number of potential issuer numbers (which is not likely to be very high). There are implementations that can do millions of hashes per second on everyday hardware, so no matter what salting you do, this is not going to be a big deal to brute force.
By sheer coincidence, when looking for something giving hash algorithm benchmarks, I found this article about storing credit card hashes, which says:
The full article is well worth a thorough read. Unfortunately, the upshot seems to be that any circumstance that makes it 'safe' to store hashed credit card numbers will also make it prohibitively expensive to search for duplicates.