Replace the `conductor_new_token` path with a token-count path

[NSagan271]

Difficulty: 🟢 Good first issue (small, well-isolated)

Scope: Small; a few dozen LoC across 3 files; mostly deletion.

Subsystems: graph · worker · conductor · node_manager_utils

Prerequisites: Read how new-token signals flow from worker to conductor. No GPU work.

Problem

The conductor_new_token field on GraphEdge
(graph/base.py) is half-deprecated. It used to drive
check_stop, but stop-checking now happens on the worker (in
_postprocess_batch's check_stop call, worker.py).
The only remaining consumer of the materialized tokens is a count, in the
conductor's handler for the worker→conductor "graphs done" message
(conductor.py):

if body.new_tokens:
    for name, tokens in body.new_tokens.items():
        pstate.num_output_tokens += len(tokens)
        for conn in request_data.streaming_connections.values():
            if conn.from_partition == partition_name and conn.edge_name == name:
                conn.token_count += len(tokens)

Yet the worker pays to materialize the actual token values on the host to
produce that list: the D→H .cpu().numpy().tolist() in _send_outputs
(worker.py), plus the buffer/flush bookkeeping
(buffer_new_tokens / flush_new_tokens in
node_manager_utils.py). We move integers to
the host only to call len() on them.

Suggested tasks

• Replace the new_tokens: dict[str, list[int]] payload on the
  worker→conductor "graphs done" message with new_token_counts: dict[str, int].
• On the worker, count tokens from tensor shapes (the new-token tensors'
  sizes are already known — see the new_token_outputs edge collection in
  node_manager_utils.py and
  _d2h_new_tokens in worker.py) instead of doing
  the D→H copy and .tolist().
• Update the conductor to consume counts directly.
• Remove the now-dead buffer/flush-new-tokens machinery if nothing else uses
  it. Consider renaming conductor_new_token → conductor_token_count (or
  similar) for clarity.

Acceptance criteria

• num_output_tokens and streaming token_count are unchanged for all existing
  models (verify against an autoregressive model like orpheus / qwen3_omni).
• No host materialization of token values occurs solely for counting — i.e.
  the .cpu() in the _send_outputs new-token path is gone. (The separate D→H
  in _prematerialize_for_check_stop is unrelated and stays; see Gotchas.)

Gotchas

• This change only touches the new-token counting path. check_stop does its own
  D→H copy in _prematerialize_for_check_stop
  (worker.py) — that's a separate part of the code that
  legitimately needs the token values on the host, so leave it alone. The goal is
  to remove the redundant materialization in _send_outputs, not all D→H
  copies of tokens.
• Confirm no other consumer reads the token values off this path (grep
  new_token, pending_new_tokens, flush_new_tokens).

New to M*? Skim How it works and the Contributing guide first.