From an EC point of view, you have one private key with one corresponding public key.

The problem is, public keys can be serialized in two ways - compressed (33 bytes) or uncompressed (65 bytes). One is slightly harder to deal with, but as storage space is more critical in Bitcoin, we prefer to use the compressed one. Thus, we now have one private key that corresponds to (from Bitcoin's point of view, as we deal with the serialized versions) two public keys. Each of these public keys has an address (as the hash is calculated from the serialized public key). So, 1 private key, 2 public keys, 2 addresses.

So when you want to import a private key, the software has to know which of the public keys (with corresponding address) should be used. The solution is to add a flag bit to the base58 encoding of the private key, notifying the importer whether or not to use the compressed public keys. Typically, these get called compressed and uncompressed private keys - but it's really just a bit saying whether the corresponding public key is compressed or not.

The only reason not to use a compressed public key is that not all software supports it (they were introduced in Bitcoind/Bitcoin-Qt 0.6 only).

@MoonShadow: EC public keys are actually not numbers but a pair of numbers (X and Y coordinate), and the Y coordinate can be calculated from the X coordinate. That is how "compression" works - it's just a somewhat less redundant encoding. Testnet is unrealted to this.