Adds generic output parser sessions, Gemma 4 reasoning channel parsing,
and complete agentic tool calling support for Gemma 4 VLM models.

## Reasoning parser

Refactors the scheduler to support pluggable per-model output parser
sessions. First implementation handles Gemma 4's reasoning channel:
strips <|channel>thought…<channel|> from streamed output and re-emits
it as <think>…</think> in reasoning_content.

- omlx/adapter/output_parser.py (new): detect_output_parser() factory
- omlx/adapter/gemma4.py (new): Gemma4OutputParserSession
- omlx/scheduler.py: threads parser session through generation loop
- omlx/utils/tokenizer.py: helpers for the session

## Tool call output parsing

_inject_tool_calling was using mlx_lm._infer_tool_parser which has no
knowledge of Gemma4's <|tool_call> marker, returning None and leaving
has_tool_calling unset so raw markup leaked into response content.

The pinned mlx-vlm (43b9b20) ships mlx_vlm.tool_parsers — a superset
that adds <|tool_call> -> gemma4 detection and a correct per-model
parser. _inject_tool_calling now prefers this, falling back to the
mlx_lm path for older installs.

## Tool result ingestion (message extractor pattern)

The Gemma 4 chat template has no handling for role=tool messages. Tool
results must appear on a model-role turn as tool_responses:

    {"role": "assistant", "tool_responses": [{"name": fn, "response": ...}]}

Passing raw role=tool messages caused the template to emit
<|tool_response> literals in content and halt after the first tool call.

Two secondary bugs: _merge_consecutive_roles was collapsing the
tool_calls and tool_responses turns into one (fixed with
_PRESERVE_BOUNDARY_KEY); _drop_void_assistant_messages was stripping
the tool_responses turn (fixed with a tool_responses guard).

Design mirrors detect_output_parser — model-specific logic stays out of
server.py:
- detect_message_extractor() in output_parser.py returns the extractor
- BatchedEngine/VLMBatchedEngine expose it as message_extractor property
- server.py uses getattr(engine, 'message_extractor', None)
- All Gemma4 logic lives in omlx/adapter/gemma4.py

## Tests

tests/test_gemma4_messages.py  - 11 tests: message conversion, tool
                                  result folding, name resolution,
                                  multi-turn agentic loops
tests/test_output_parser.py    - reasoning parser session
tests/test_scheduler.py        - scheduler integration
tests/test_utils_tokenizer.py  - tokenizer helpers