I know that installing QVS on a virtual machine is not the best thing to do, but I'd like to know if there are rules to follow to dimension correctly a QVS installation for test purposes, remembering that I have to manage the same numbero of records like in production.
You only need to be concerned about virtualization in a production environment. Setting up test servers in a virtualized environment is a common thing. Can you not make the test server the same size as the production server? Are you testing function or performance?
There are no numbers that can be even remotely applied to any scenario where only the qvw document size and hardware specs are known. The performance and load of a given QVS is always something that depends on the data model (unique values, table scheme, data amounts, data types), object definitions (chart/object expressions, set analysis, alternate states, section access), user behavior (number of users, click frequency, usage of data) and data refresh rates (reloads, data exchange/purge rate) and so on.
You can make a comparision to tuning an SQL Server; you never know how the server will behave until you have users actually querying the data. And until then, you really don't know how to construct your data indexes. It's no difference in a QlikView environment.
So, unfortunately, the only way is to test scenarios and tune after results.