Don't wait for Phabricator events in the 'events' service, wait through a Taskcluster hook instead

[marco-c]

Taskcluster now supports triggering a webhook through a unique url (with a token). We could build a new hook that is triggered by each Phabricator diff (same config as currently), creating a task associated to the diff. That task would run the same workflow as events:

• Clone base repo, using a Taskcluster cache and robustcheckout
• Apply the diff on tip
• Push to try
• Update phabricator harbormaster state for the diff

This would resolve the issues around disk space that we have on Heroku, allowing us to do things like #2004 and #2204. This would also allow us to make the 'events' service slimmer. If we do both this and #945, we could get rid of the 'events' service altogether.

[La0]

The overall workflow would be :

1. Phabricator triggers a webhook on Taskcluster

   • using a new hook declared in fxci-config repo
   • using Taskcluster recent public webhook trigger

2. The hook triggers a new task with the request from the caller (in payload

   {
   firedBy: "triggerHookWithToken",
   taskId: 'IgfFQSAqQwysozyeB7udBw',
   payload: {..}               // API call payload
   }

   We know from our libmozevent implementation that Phabricator is able to send

   • diff ID (that's the only required info)
   • target PHID (required to send messages back to Phabricator
   • repo PHID
   • revision ID

3. From there we can reuse code from libmozevent, without using the Bus (used on events to run continously). Internal wokflow is the one mentionned by @marco-c above:

   • clone base repo using TC cache & robustcheckout (implemented in libmozevent)
   • apply the patch (also in libmozevent , just requires to create a PhabricatorBuild to match current implem)
   • push to try is immediate
   • updating Harbormaster state is also available in libmozevent

I suggest we implement this in a new sub-project in this repo, named trigger. It would depend on libmozevent and only implement glue-code to retrieve infos from the Taskcluster payload (or from local CLI arguments to debug/develop)