fix: harden FBHTTPServer/FBTCPSocket against races and protocol gaps

[mykola-mokhnach]

Follow-up fixes for the Network.framework HTTP server (#1221) found while porting it to a downstream WebDriverAgent fork:

• Defer FBWebServer's shutdown delegate call so the /wda/shutdown response is actually written before the socket/connections are torn down.
• Cancel FBTCPSocket connections on the socket queue during -stop so an already-issued write isn't raced by the cancellation.
• Cancel the listener on startup timeout/failure instead of just dropping the reference, and surface POSIX errors (e.g. EADDRINUSE) undecorated so FBWebServer's port-retry check can match them.
• Stop accumulating received data via repeated mutableCopy+append (O(n^2)); append into a single NSMutableData instead.
• Cache parsed request headers per connection while the body streams in, instead of re-scanning/re-parsing the whole buffered header block on every incoming chunk.
• Remove a client's buffered state before responding with a terminal error (413/400/501) so a still-streaming body can't re-trigger the same response.
• Reject Transfer-Encoding: chunked explicitly (501) instead of silently treating the request as bodyless and desyncing the connection.
• Apply default headers (CORS, Server) to error responses, not just successfully-routed ones.
• Give multiple "*" wildcard segments in one route path distinct params keys instead of overwriting a single "wildcards" key.
• Return valid WebDriver protocol error JSON for 400/404/413/501 responses instead of raw strings, matching the shape route handlers already use.

[eglitise]

Looks like there are some merge conflicts

[mykola-mokhnach]

Looks like there are some merge conflicts

conflict resolved