An update on my investigation if anyone out there is having the same problem.
As the cat/proc/interrupts showed, we are sharing IRQ11 for the uhci_hcd, et0 and our infamous nvidia card.
As I pointed out, when I turn on my multi-head system, it goes blocking from seconds to seconds, untill the network card receives some traffic. This is showing that the system is only getting stable AFTER THE CARD IS USED, or in other words when the card sends an IRQ11.
I don't know much about Linux USB, but since the uhci_hcd is also in IRQ11 I tried to make it throw and IRQ11 by inserting a memory key, but to no avail.
Now, after supreme hours of supreme suffering, I discovered that a good'ol "service network restart" (Fedora Core 3, btw) would make the network card make it's IRQ11 and the system works.
So what we can see from all this is:
1) The nvidia driver is expecting some kind of action from the IRQ11 card(s) and is aborting some kind of loop after it receives an IRQ11 generated from a device that isn't hers (the NIC).
2) We can make the NIC send an IRQ11 and make the nvidia driver come back to the real world.
3) This is not such a strange problem after all... there goes my Guines!!