Best practice for deduplicating in-flight requests

### Motivation

Hello,
I wanted to ask how to approach with removing duplicated request in flight? I was wondering about creating dedicated Middleware that will catch all simultaneous calls, fire first one and on return share result with other callers. This might work fine, but I faced an issue where the first call is cancelled. This triggers cancel handler in URLSessionTransport that cancels request and returns nothing. This propagates to all callers waiting for the response.

### Proposed solution

In iOS 27 there will be shield mechanism https://developer.apple.com/documentation/swift/withtaskcancellationshield(operation:)
but I need to have compatibility with older iOS versions.

I thought about shared detached Task, but ClientMiddleware protocol assumes that next closure is non escaping, which would be required in order to pass it into a new task.

Other option I would see is to be able to control cancel behaviour in URLSessionTransport. A flag in configuration that determines if canceling task means that dataTask is cancelled/ operation checks for cancellation. Right now call is fired anyway by URLSession, but later checks in e. g. bidirectionalStreamingRequest:
```
try Task.checkCancellation()
return (try HTTPResponse(response), responseBody)
```
makes impossible to react in higher layers. I would like to have more freedom of controlling such behaviour. If task is cancelled, it would make sense to still have a response and its body to evaluate and maybe clean up.

### Alternatives considered

_No response_

### Additional information

_No response_

Start by reading the ClientMiddleware protocol and the cancellation path in URLSessionTransport, especially the behavior around bidirectionalStreamingRequest and Task.checkCancellation(). Compare the proposed cancellation-control and shared-task approaches with the existing nonescaping next closure. The issue does not define an accepted design or a specific test that would establish when the work is done.

Written by the indexing model from the issue text.