fix(chat): make get_tokens() robust to non-alternating turn histories

[cpsievert]

Ports tidyverse/ellmer#1157 (fixes tidyverse/ellmer#1131).

Problem

Chat.get_tokens() computed per-turn token counts by pairing user and assistant turns positionally, which only works when turns strictly alternate. It raised ValueError for histories that don't:

• a trailing user turn with no completed assistant response yet — this arises in normal use when a request fails after the user turn was recorded, or after add_turn()
• consecutive assistant turns (possible via set_turns())
• a lone user turn

Fix

The computation is now driven off complete assistant turns instead of turn positions. Each assistant turn's input token count is cumulative over everything sent so far, so the implied user tokens for a round trip is simply the difference between consecutive requests — no alternation assumption needed. Histories with no complete assistant turns now return an empty list, and output for well-formed histories (including tool-calling loops, where tool results ride in user turns) is unchanged.

The only remaining ValueError is for assistant turns that genuinely lack token counts (e.g. fabricated via set_turns()), pointing at .get_turns() for raw values as before.

Tests

New credential-free tests cover: trailing user turn, lone user turn, consecutive assistant turns, tool loops, and the missing-token-count error. Existing test_tokens_method expectations are unchanged.