Support for alternate transports

What would it take to substitute in an alternate transport between HttpClient and Kestrel for fast local machine proxying? E.g. can we replace Sockets with Named Pipes? How much faster would that be?

- Sockets
- ~~Named Pipes~~
- Shared Memory
- Unix Domain Sockets

Start by tracing the HttpClient-to-Kestrel path and the current Sockets transport, then compare Named Pipes, Shared Memory, and Unix Domain Sockets. Done means producing a measured feasibility and performance comparison and a scoped decision about whether an alternate transport should be supported.

Written by the indexing model from the issue text.