feat: expose logical conversation rounds over flat turn history

[cpsievert]

Why this matters

Chatlas stores a flat sequence of SystemTurn, UserTurn, and AssistantTurn objects. A single user request that invokes tools can span several user/assistant turns, making transcript review, UIs, and tracing consumers reconstruct the logical exchange themselves.

Ellmer added a read-only Round abstraction in tidyverse/ellmer#1056 (59bbe9d4): each round starts with input turn(s) and includes the assistant/tool-result turns that follow. Storage remains flat.

Scope

• Design and add a public, read-only Python API that groups current history into logical conversation rounds without changing how turns are stored or submitted to providers.
• Include a clear definition for tool loops, system turns, and partial assistant turns.
• Provide an ergonomic way to retrieve all rounds and the final round.
• Ensure returned round objects preserve the original turns rather than duplicating provider state.
• Add tests for ordinary exchanges, multiple tool-result cycles, system-prompt inclusion, manually supplied history where supported, and interrupted/partial responses.

Starting points

• chatlas/_chat.py: get_turns(), get_last_turn(), and partial-turn helpers.
• chatlas/_turn.py: SystemTurn, UserTurn, AssistantTurn, and partial_reason.
• Chat.get_turns(tool_result_role="assistant") already offers display-oriented collapsing, but it is not a round API.
• Ellmer reference: git -C ../ellmer show 59bbe9d4.

Non-goals

• Do not replace the flat internal history or alter provider serialization.
• Do not add mutation APIs such as set_rounds() without a separate design decision.

Design note

This is a public API addition. Settle names, object shape, and system-turn semantics before implementation; the ellmer behavior is a useful reference, not a requirement to mirror verbatim.