feat(server): publish [:grpc, :server, :rpc, :abort] telemetry

[ziomecka]

An RPC that the adapter stops before it returns publishes no server telemetry. The exit signal that stops the RPC process does not unwind it, so the :telemetry.span/3 around the call (telemetry.ex:95) publishes neither :stop nor :exception. Expired deadlines, client cancellations and dropped connections all leave no record.

This adds [:grpc, :server, :rpc, :abort], published from the cowboy handler process, which outlives the RPC process.

• Publishes the event just before the exit signal, at both abort sites: send_error/4 (handler.ex:466) and terminate/3 (handler.ex:528). Both now go through one abort_rpc/2.
• Metadata is :stream, :server, :endpoint, :path, :pid, :reason, with :duration measured from the arrival of the request. There is no :function_name: it is resolved in the RPC process.
• The loop state gains stream, endpoint, route and started_at.
• The event is guarded to one per call. Both sites can run for the same call, and Process.exit/2 is asynchronous, so Process.alive?/1 can still be true at the second one.
• grpc_core is documentation only. The event name is built from the public GRPC.Telemetry.server_rpc_prefix/0, so grpc_server still compiles against the released grpc_core.

Measured on a service behind a proxy enforcing a 2400 ms deadline, over 6 hours at ~800 req/s: the longest server span recorded for the hot method was 2386 ms and none was above 2400 ms, against ~40 ms normally. Every call that crossed the deadline left no server-side record.

Tests in handler_test.exs:

• An expired deadline publishes :abort, and neither :stop nor :exception.
• A dropped connection publishes :abort, covering the terminate/3 site.
• One :abort per call when both sites run, using a handler that traps exits so the RPC process is still alive at terminate/3.
• A call that returns publishes :stop and no :abort, since terminate/3 runs for every request.

Three of the four fail with the publish call removed. Full suite green, mix format --check-formatted and MIX_ENV=test mix compile --warnings-as-errors clean, CHANGELOG entry under Unreleased. Verified on Elixir 1.20.3 / OTP 28 only; CI will be the first run of the older matrix rows.

Two choices I am glad to change: the one-per-call guard uses the process dictionary, and threading it through the loop state instead means changing send_error/4 to return {req, state} and updating its seven call sites; and :abort is a guess at your preferred name.