Strictly enforce canonicaljson requirements in a new room version

[clokep]

This PR does the following:

• Adds an experimental room version which strictly adheres to the canonical JSON specification:
  • Rejects integers outside of the range of [-2 ^ 53 + 1, 2 ^ 53 - 1]
  • Rejects floats
  • Does not parse Python special values (NaN, Infinity, -Infinity)
• Logic to strictly enforce this is added to the client-server API and the federation API.

Replaces #7356

Implements MSC2540: matrix-org/matrix-spec-proposals#2540.

To Do

• Perform more testing with federation.
• Does this need to be optimized? (Some really rough numbers show that this about doubles the runtime of event_from_pdu_json.)
  • No!
• Can we consolidate some of this validation logic further?

[clokep]

Requesting review on this to get some thoughts on the approach.

[clokep]

I did some testing with two different clients:

• The server rejects bad messages from a client.
• The server will reject a bad message over fedation from another server.
• The server will reject a bad message while backfilling from another server.

Are there other scenarios to test?

[ara4n]

It's not a canonicaljson thing, but, while we're adding a new room version, might be good to check that power levels have to be ints rather than strings (and perhaps sanitycheck other type safety in synapse)

[richvdh]

power levels have to be ints rather than strings (and perhaps sanitycheck other type safety in synapse)

that is the thin end of a very large wedge. I think we're better off keeping this constrained for now.

[anoadragon453]

CanonicalJSON indeed does aim to represent all floats as ints, just in case anyone was unsure about blocking all floats: https://matrix.org/docs/spec/appendices#canonical-json

[clokep]

Correct, as the description says all floats need to be rejected. I didn't find the specification super clear here unless you read into the grammar though. I think we could improve that.

[richvdh]

this looks like generally the right plan to me.

I wouldn't be too worried about performance here: event_from_pdu_json is called only once for each event, and it tends to happen on workers rather than on the master.

[richvdh]

lgtm