[javier-godoy]

Clients that hold Docker API streams open, like events, followed logs, wait and exec/attach, don't work well behind the proxy as it stands:

• Buffering delays streamed responses. nginx buffers proxied requests and responses by default, so output from events, logs, image pulls, builds and exports can reach the client late.
• Large uploads are rejected. The default client_max_body_size of 1 MB makes uploads like build contexts and docker load fail with 413.
• Idle streams are cut off. proxy_read_timeout closes any upstream connection that is quiet for 60 seconds, which ends idle events subscriptions and followed logs.
• The connection limit is fixed. The generated config hard-codes worker_connections 1024. Since it runs a single worker, that is the limit for the whole proxy. Once long-lived streams from many clients fill it, nginx refuses new requests with worker_connections are not enough, and there's currently no way to raise the limit.

This PR:

• Turns off request/response buffering and the body size limit, so streams pass straight through and uploads of any size are accepted.
• Adds WORKER_CONNECTIONS to set worker_connections. It still defaults to 1024.
• Adds PROXY_READ_TIMEOUT to set proxy_read_timeout, for example 24h for long-lived idle streams.
• Adds TCP_KEEPALIVE to set so_keepalive on the listening port, so the connections of clients that have disappeared are detected and closed. This is useful together with a long PROXY_READ_TIMEOUT.

The new settings are only written to the config when set, so nginx's defaults stay in place otherwise. All of them are documented in a new "nginx Settings" section of the README.

Each commit is self-contained, so I'm happy to split this into separate PRs if you prefer.