Streaming /v1/messages: message_start reports input_tokens 0 and no cache fields #93

Streamed /v1/messages sends message_start with usage = {"input_tokens": 0, "output_tokens": 0} and no cache fields, also with --enable-cache-report. The real prompt and cache counts arrive only in the final message_delta. Clients that read input usage from message_start see 0 on every streamed turn.

Upstream: FlashML-org#718 (repro, cause and proposed fix). The repro there gives the same output on 5080deploy c23d2d5:

cache_report=True message_start: {'input_tokens': 0, 'output_tokens': 0}
cache_report=True message_delta: {'input_tokens': 176, 'output_tokens': 5, 'cache_read_input_tokens': 1024}

Proposed: emit a prompt-admitted event from _generate_events_impl when the admission ack arrives, and send message_start with the real usage on it.