fix(perplexity): keep tool calls when merging consecutive assistant messages

[BlueX888]

PerplexityLLMAdapter merges consecutive same-role messages so the request satisfies Perplexity's strict role alternation. The merge moved only content off the message it absorbed before dropping it, so a message carrying tool_calls lost them — and two consecutive tool messages collapsed into one that answers only the first call.

Why

In step 2 of _transform_messages (src/pipecat/adapters/services/perplexity_adapter.py:151), any two adjacent same-role messages are merged: next_msg["content"] is extended into current (:158-161), then next_msg is popped (:162). Nothing carries tool_calls across the merge, and nothing treats a tool message specially — yet its tool_call_id is the only thing tying its result to a call.

A history in OpenAI message shape therefore comes out of the adapter wrong in two ways:

• [assistant(text), assistant(tool_calls: c1), tool(c1)] -> [assistant(text), tool(c1)]. The tool message answers a call that is no longer in the request.
• [assistant(tool_calls: c1, c2), tool(c1), tool(c2)] -> [assistant(tool_calls: c1, c2), tool(c1)]. Both results land on c1, and c2 is left unanswered.

Changes

• The merge carries tool_calls from the absorbed message onto the surviving one, and moves content into it when the surviving message has none — the call can precede the text it was made alongside.
• Consecutive tool messages are left as they are, since merging them would leave both results on the first call's id.

Testing

On the tree with the source change reverted, the three new tests fail:

FAILED ...::test_merging_assistants_keeps_tool_calls            # KeyError: 'tool_calls'
FAILED ...::test_consecutive_tool_messages_keep_their_own_ids   # AssertionError: 4 != 5
FAILED ...::test_merging_tool_call_only_assistant_keeps_content # KeyError: 'content'
===================== 3 failed, 13 passed in 0.68s ===================

With the change:

$ uv run pytest tests/test_get_llm_invocation_params.py::TestPerplexityGetLLMInvocationParams
16 passed

$ uv run pytest tests/test_get_llm_invocation_params.py
179 passed

$ uv run pytest tests/test_get_llm_invocation_params.py tests/test_openai_compatible_token_usage.py
228 passed

$ uv run ruff check            -> All checks passed!
$ uv run ruff format --check   -> 1412 files already formatted
$ uv run pyright               -> no findings in the changed file

Notes

No issue to reference; this is not from a report. I left changelog/ alone because the fragment filename is the PR number — say the word and I'll add <this PR>.fixed.md.