The first step is to choose how you will mix individual disks using disk striping (RAID 0) to get the best performance. Each disk doesn't have great throughput but if you stripe a few disks together they you get much more throughput.
When you create a disk you can choose the whether to have caching on or not. There are two sets of throttles one for cached and one for uncached disks so creating a volume out of cached or uncached you can get much higher throughput
In most instances, you can't change the fact that the local SSD comes from the cache drive so that will take your throughput if you use it (you should use it, it is very fast)
If you need more throughput than the cached bucket allows then you can use uncached disks to increase the throughput but there is no caching so it will be slower
If you have a cached disk, a cache is on a different part of the local ssd that you can't see. If requests are served from cache then request does not count towards your throughout so you can potentially get even higher than the max but if you are designing an IO system on azure VM's I would make sure you're happy with the performance when nothing is in the cache.
If you push your disks then it is the max throughput before you start getting throttled, you know your application though and will know if you need all the volumes at the same time or maybe you can balance things so you can have different disks busy at different times.
Well, it isn't the end of the world, IO requests will take longer if you get throttled for 10 seconds once a day, who cares? if you get throttled every second of every day then maybe do something about it if your app is slow etc.