Cache jwks

[mzaniolo]

On #47 we added the option to fetch the public keys from the IDP. Now we have to cache the jwks to avoid fetching it on every request. We should look for the cache control header to determine for how long the cache should be kept

[sirewix]

From the business logic perspective I think it's a pretty straightforward, we just check if stored jwk is older than cache-control's max-age then we renew it.

From the technical perspective I suppose we need some kind of mutex/rwlock

[sirewix]

Ok, I did some digging and it's not so straightforward. For the caching part itself it'd be nice to use locking to avoid extra calls, but looks like it's quite tricky in python. But since python is single threaded we can just "write once" and it should not cause much issues except that under a heavy load when the data expires, many threads would fetch the data at the same time, parse it and "save once".

However the bigger issue is that synapse.http.client.get_raw returns the body of the request but not the headers, and we could try copy paste and write our own implementation of get_raw but it calls make_deferred_yieldable inside it and we cannot really import anything from synapse module directly and it has to be taken from ModuleApi, which doesn't seem to provide this function.

There's also synapse.http.client.get_file that does return the response headers, but it needs output_stream: BinaryIO as input and it looks like it can only save the result as a file.