How many ‘subscribe’ events should be collected before running a Facebook ad optimized for subscribe?

While this is counter-intuitive, against recommended best practices and may be temporary, I have heard from FB technical employees that the optimization is always better when using the "Purchase" event of Facebook, regardless of what you mapped onto it. I have optimized for trials successfully by mapping it the FB "Purchase" event.

In your case, if you have enough (real) Purchase events to optimize for, why would you want to optimize for the trials in the first place?

I have done a bit of testing around this. Optimizing towards a free trial was worked best for a couple of our apps. Big thing to make sure of is that your upsell screen matches your optimization. We found a little bit of success optimizing towards revenue with an upsell screen that focused on non free trial plans. We were not able to scale it more than a couple hundred dollars a day, but the market was already pretty saturated for that product. Could have been effective if tested earlier in the lifecycle.

Or it is just the case that FB ad works well with Purchase optimization but not with Subscribe optimization?

Yes, this is generally true: fewer apps implement a subscription mechanic than a purchase mechanic, so Facebook has a larger set of purchase events from which to draw when optimizing an audience for ad targeting. 

In general, I think AEO works best on the purchase event (vs. every other event) for this reason. If you want to optimize against Subscribers, it might be best to implement an LaL workflow using seed lists of subscribers.