Improve import stats performance by emitting start and end together

[op]

Decrease lock contention on imports by sending start and end events all at once.

[danielnorberg]

With the current internal telemetry collection implementation emitting start and stop events at the same time would cause the resulting spans to become zero length, although maybe that's something that can be fixed.

[dflemstr]

Maybe out of scope for this particular PR but I wonder how we can attack the main bottleneck of this? If the main issue is contending for the queue, this PR reduces the overhead by a constant factor which is great but I guess the contention is still O(n) in terms of something (nr of threads)?

[danielnorberg]

I would be surprised if contention is the problem given that python is mostly single threaded. It might be good to benchmark and measure to get a sense of where the bottleneck is.