By the way, SatoshiDice is (afaik) not using compressed pubkeys. That's because bitcoinj doesn't use them. SatoshiDice is open source, it uses a fork of the library. We can probably halve the amount of data it generates just by fixing that.
It only saves you 32 bytes per txin, so more something like 20%. That said, yes, that would be a very welcome change.
It's on the order of a couple hundred megs. I forgot the exact size. Perfectly feasible to hold entirely in RAM for now.
In the format used by the current git head of the reference client (which will become 0.8 ), it's about 122 MiB now. Together with indexes and other overhead added by LevelDB, it's 143 MiB. Unpacked in RAM it's probably close to 5x as much (though there is little reason to have the entire database unpacked in memory).