Pool CancellationTokenSources

[MihaZupan]

Fixes #902
Contributes to #1055

I will post an allocation profile here before merging as well.
It should be the same as #876 (comment) for 6.0 + a few state machine/object size reductions.

load	lin-citrine-net60-yarp-http-http-100B	lin-citrine-net60-yarp-http-http-100B-local
Requests	14.591.185	14.848.700	+1,76%
Bad responses	0	0
Mean latency (us)	1.050	1.032	-1,74%
Max latency (us)	48.932	39.009	-20,28%
Requests/sec	243.194	247.529	+1,78%
Requests/sec (max)	262.924	267.700	+1,82%

The RPS change across multiple runs seems to be around 1 - 1.5%.

This change relies on #1271 for correctness - a rogue StreamCopier must not be allowed to run when we're exiting SendAsync.

[Kahbazi]

CancellationTokenSource in ActiveHealthCheckMonitor could also be pooled.
https://github.com/microsoft/reverse-proxy/blob/762b1a43216cb2161a0dcb422036bb6fb9e84351/src/ReverseProxy/Health/ActiveHealthCheckMonitor.cs#L123

[MihaZupan]

Looking at the ActiveHealthCheckMonitor logic, I'm seeing an unrelated bug (#1299), so I will look at using a pooled CTS there in a separate PR.

[MihaZupan]

As expected, the allocations for CTS machinery are saved per request (4 objects - 304 bytes).
StreamCopyHttpContent, HttpForwarder.SendAsync and StreamCopier.CopyAsync's state machines are 8 bytes smaller.

[Tratcher]

I'm amused we're taking turns modifying the same lines of code.