add renderes clear thinking

[samsja]

Note

Medium Risk
Changes Nemotron3 prompt rendering semantics for reasoning_content by introducing a clear_thinking flag and a new registry entry to keep historical thinking, which can alter tokenization/training masks and downstream model behavior for Nemotron3 runs.

Overview
Adds a clear_thinking option to Nemotron3Renderer that controls whether reasoning_content in earlier assistant turns is collapsed to an empty <think></think> block or preserved, updating the assistant rendering path accordingly.

Registers a new nemotron3_keep_thinking renderer variant (subclassing Nemotron3Renderer with clear_thinking=False) in renderers/base.py so callers can opt into keeping thinking content across the full conversation history.

^{Reviewed by Cursor Bugbot for commit 12a9fd3. Bugbot is set up for automated code reviews on this repo. Configure here.}