server: tool role detection and create checkpoints only on user messages

[Kononnable]

Overview

Changes in #24176 introduced user message detection. There was a problem with tool call responses which were classified as user messages. Alternative approach (#24176 (comment)) was chosen to create checkpoints on user and tool response messages (guarded by a larger checkpoint_min_step. However, this approach has a problem of not utilizing full batch size during prefill of old conversation with many small tool calls.
This PR introduces detection of tool call response to autoparser and reverts the state to original idea - creating checkpoints only on user messages (no tool responses).

cc: @aldehir, @ggerganov

Prior work: #24176
Fixes: #25213
Fixes: #25320
Fixes: #25023

Additional information

Problem with recognizing tool responses is that they're send as a user message, so they're always caught by a user_msg_start. We can check if tokens are caught by multiple delimiters and assign it to the longest one (most specialized).

--- Reasoning & Content Structure ---
0.06.366.668 W user_msg_start: <|im_start|>user
0.06.366.668 W assistant_msg_start: <|im_start|>assistant
0.06.366.668 W tool_response_start: <|im_start|>user
<tool_response>
0.06.366.669 W reasoning_mode: TAG_BASED

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[aldehir]

I have another patch to fix the prefill issue. Will submit later today.

Detecting the tool responses still has value but that's at @pwilkin's discretion.

[Kononnable]

Closing as related issues were fixed using current approach (is_last_user_message approach).

If detecting tool responses would be useful I can create a PR with just those changes (plus necessary changes not to break current checkpoints behavior). However, adding it feels like adding a dead code, unless this detection is required by some already planned feature.