Event send time graph is misleading due to including errors in the graph

This issue has been migrated from [#13854](https://github.com/matrix-org/synapse/issues/13854).

---

### Description

Blocking a high traffic room meant that the event send time dropped to <10ms on average, which is not physically possible with the infrastructure I have in place - errors shouldn't count towards the "Event Send Time" graph/metric.

### Steps to reproduce

- Get lots of 403 or similar errors in sending events
- See that your send times look great on the graph now

### Homeserver

t2bot.io

### Synapse Version

1.67.0

### Installation Method

_No response_

### Platform

Ubuntu server, postgresql on a different box.

### Relevant log output

```shell
N/A
```

### Anything else that would be useful to know?

_No response_

Start by locating the Synapse code that produces the “Event Send Time” graph or metric, then reproduce the reported case with many 403 errors in event sending. Done means failed event sends no longer make the displayed average send time appear artificially low.

Written by the indexing model from the issue text.