What fillfactor for caching table?

The key words here are:

1. "heavily updated"
2. "in the table for 2-3 hours".

Point 1. is indication for a lower fill factor, while 2. is the opposite. It helps performance if multiple row versions are stored on the same data page. H.O.T. updates would achieve that. Read here or here. They need some wiggle room on the data page - like dead tuples or space reserved by a fillfactor < 100. But they can only do their thing, if no index involves any of the updated columns, which should be true for your case.

Another important factor here would be the tuple size (in comparison to your page size (which is most commonly 8 kb). More details in this related answer:

• Measure the size of a PostgreSQL table row

If the tuple size is 4 kb or more, reducing the fill factor would be futile, since there can never be more than one tuple on a data page. You might as well leave it at 100 (which is the default anyway). However, some data types are "toasted" and stored out-of-line if they exceed a size limit, so tuples requiring that much in the main relation fork are rare.

Whatever you do, VACUUM will be run often. And that's generally a good thing, I wouldn't worry about that. You create lots of dead tuples. VACUUM identifies dead rows that are not visible to any open transaction any more. The manual:

The standard form of VACUUM removes dead row versions in tables and indexes and marks the space available for future reuse.

Bold emphasis mine.
You can play with per-table settings for autovacuum to trigger it less (or more) often for this table only:

The default thresholds and scale factors are taken from postgresql.conf, but it is possible to override them on a table-by-table basis;

Bold emphasis mine. In particular with autovacuum_vacuum_threshold and autovacuum_vacuum_scale_factor. Running VACUUM a lot might actually be a good idea, instead of a very low fillfacter. That depends on access patterns. If all tuples live, say, 3 hours and each is updated several times, I would still lower the fillfactor to something like 50. You'll have to test and find the sweet spot.

Alternatives

All this aside, since your data seem to be volatile to begin with: use an UNLOGGED table:

Data written to unlogged tables is not written to the write-ahead log (see Chapter 29), which makes them considerably faster than ordinary tables. However, they are not crash-safe: an unlogged table is automatically truncated after a crash or unclean shutdown. The contents of an unlogged table are also not replicated to standby servers.

Bold emphasis mine. Don't use this if your server might crash and you still need the data afterwards. But if we are talking about session data for web applications, this might be an acceptable price to pay.

Or, even more radical: Use a key-value store like Redis if you can do without the features and security provided by an RDBMS altogether.