I thought I'd chip in as I happen to work with the OP and we discussed the project today.
The OP is correct that he needs a tonne of light for this project. LEDs of the sort he listed are what are currently used for experiments of this sort. In related experiments in vertebrate brains, the light source is usually a laser. The light-sensitive channels in question are inside the fly and light has to get through the pigmented cuticle. Also, these channels are inserted into regular neurons using genetically modified flies. The photoreceptors in your eye have molecular signalling cascade that amplifies the response to light; these regular neurons do not have this and they don't express the channels at high density. You have to make up for this by dumping a load of light at the flies.
Based on previous work, we estimate that we need a light source which delivers a maximum of about 1 mW/mm2 at the vial which contains the flies. An effect should probably be observable starting at intensities of about 0.1 to 0.2 mW/mm2. To measure the output you really need a light meter located at some known and reasonable distance from the source. You can't go from the power used because LEDs aren't 100% efficient. You can't go from the LED size (or the emitter size) alone because the flies aren't at the emitter. Neither can you increase light levels much by adding more LEDs. More LEDs will increase light uniformity and you'll get the sample closer to the radiant flux at the source, but you'll never get beyond that flux by adding more LEDs.
I measured the LED's output with a light power meter at a distance of about 1 cm. The LEDs can reach the required maximum light level when driven at about 500 to 600 mA. 1 mW/mm2 is a lot brighter than you might think. It's unpleasantly bright when viewed in an indoor environment. Titrating the light levels and exploring various light-delivery regimes are experiments which need to be done; but you really want a system that's capable of delivering more than enough light because if your experiment doesn't work then you need to be able to discount low light as a possible explanation.
We have chosen to set up 6 LEDs per vial for a total of 4 vials. We're planning on driving the LEDs in series using a constant current driver (BuckPuck DC LED Drivers). We would drive the LEDs in two batches of 12, with one driver for each batch. Each batch will therefore need a 30V supply. The BuckPuck LED drivers we're planning on using have a 0-5V analog input which can be used to either set light intensity or flash the LEDs at the desired rate. We are planning on using a colleague's stand-alone function generator (GitHub - open-ephys/GUI: Archived source code for the Open Ephys GUI) for this task, as we want to easily explore a variety of light presentation regimes. Obviously once an optimal presentation regime is chosen then we could substitute the function generator something cheaper. e.g. just set up an Arduino that delivers the optimal waveform.