The point is to provide entropy. Truecrypt must generate a secret key for the volume. It does so by generating a bunch of random bits. Here (as often in cryptography), what is important is not really that the bits that make up the key are random in a statistical sense, but rather than the key cannot be predicted or reproduced by an attacker. A computer is a deterministic machine¹: if the attacker knows in what state it was in when you started to generate the key, he can run Truecrypt and generate the same key.
By moving the mouse, you are providing input that the attacker cannot reproduce. The more input you provide, the harder the key will be to reproduce. For example, if the computer only recorded a single motion as left or right, then there would only be two possible keys, and the attacker could try them both; the key would only have 1 bit of entropy (no matter how long the key is). Ideally, the key must be completely random; if the key is, say, a 128-bit key, the random number generator must have 128 bits of entropy available. Human movements are somewhat predictable (you aren't going to move the mouse two meters left), but the more you move, the more entropy you feed into the pool.
The mouse motion is not related to the 100,000 rounds. The rounds are a different issue, related to how hard it is for the attacker to reproduce your password. Humans are notoriously poor at choosing and remembering complex passwords, so the attacker can try all plausible passwords by brute force. For this reason, cryptographic systems that use passwords don't use them as-is, but perform some computation (a cryptographic hash, say; PBKDF2 is generally recommended these days) on the password many times over. This computation is expensive; its running time is proportional to the number of rounds. The system must perform this iterated computation once per password attempt; the attacker must also perform it once per password attempt. If it takes 1 second for your system to process your password when you mount the volume instead of 10 microseconds, it's not a big deal, because password processing is only a tiny fraction of what you use your CPU for anyway. But for the attacker, who's spending all his CPU time brute-forcing passwords, being able to perform only 1 cracking attempt per second and not 100,000 per CPU is a big hit.
¹ Some computers have a hardware random number generator, which derives its randomness from physically impossible to predict (or at least hidden and extremely hard to predict) sources. Nuclear decay is good for this but impractical. On mobile devices, camera white noise works fairly well. But many computers lack such a hardware random generator.