It's not OS's your students need experience with. It's software and programming languages they need experience with. If they're going to go into experimental science, they would also benefit from building hardware and interfacing it with their computers. (i.e. Some basic electronics)
I do experimental quantum physics work in a university. We use everything. OSX, XP, Windows Server (no Windows 7 or Vista installs surprisingly!), and a few distros of Linux. Sometimes we are forced to use a specific OS (usually Windows) because some piece of equipment we're interfacing with only has drivers for one OS. When that isn't the case, it's personal preference. (I gravitate towards Linux distros with decent KDE environments.) Really, you shouldn't worry about what OS your students use. Ideally, give them a chance to try out a variety of OS's.
The applications are what's really important and the big ones tend to be mostly the same across platforms. If you're doing basic (or not-so-basic) simulations or analysis, you're probably going to use Matlab or Mathematica. Something requiring higher performance will probably be written in a low level language like Fortran. (Yes, Fortran. It's surprisingly good for Physics work. Try doing linear algebra in Java or C and you'll just waste a lot of time writing tools.) If you're running an experiment you might do a little driver work in C or C++, but odds are you'll tie things together with something like Labview. Origin also gets used a fair bit for plotting and curve fitting even though it has a pretty horrible interface. Excel, gnumeric, etc. just aren't as good at fitting. For writing papers it's Latex and nothing else. Many people use Latex add-ons like beemer to make presentations as well instead of powerpoint. I'm sure other people can suggest software to get, but it's going to get expensive fast unless you can get some free educational samples, which you should probably try asking for. You might be surprised by what you get!
Here's my ideal environment for your students:
-Not one OS, but many. They should be exposed to something they don't use at home. This will help them become adaptable.
-These OS's should not be locked down. Locking them down will stifle your students ability to learn. Heck, encourage them to try breaking and fixing things. You should probably, however, create disk images so you can easily restore the machines to a useable state if they are wratched. Your sysadmin will hate this idea and would probably prefer to lock things down tightly. Just remember that if sysadmins had their way nobody would ever use their systems.
-Get as much far-out scientific software as you can. Let your students play with it. Encourage them to try checking their Calculus assignments with Mathematica or Matlab, or perhaps write them up in Latex.
-Get some hardware to hook up to the computers. Find basic sensors like thermocouples or photo-diodes. Get some USB-interface chips, prototyping breadboards, and misc components and put your students to work interfacing those sensors with a computer. They might find it impossible, or they might surprise you. Being able to tackle tasks they're not prepared for with minimal guidance is one of the most useful skills you can teach them.
-Don't make boring lessons like, "Today we're going to learn how to print, "Hello World" in Java!". Give your students projects. Ambitious projects. The sort you don't know how to do. Give them lots of class time to work on it. Even if they're doing stuff you don't know anything about, talk to them about it. Ask them what they've done, what their current problems are, and what they plan to do. It's their problem to solve, but you're the coach who helps keep them on track.
If you do even a fraction of the above, your students will be well ahead of 99% of the students coming into University.