Add opt-in recurrent profiling instrumentation

[xxxkkw]

Summary

• Add opt-in recurrent profiling instrumentation behind MLX_LM_PROFILE_RECURRENT.
• Emit structured JSON timing records for recurrent paths including Mamba, SSM, gated-delta, and RecurrentGemma code paths.
• Keep default behavior unchanged when the environment variable is unset.

Environment

• macOS Darwin 25.4.0 arm64
• Apple M1 Max, 32 GB unified memory
• MLX / mlx-lm virtual environment on Apple Silicon

Real-weight benchmark

Model: mlx-community/mamba2-130m-4bit with a config override for intermediate_size=1536, matching the model config's expand * hidden_size.

Protocol: 100K input tokens and 10K generated tokens, prefill_step_size=2048, greedy generation. This model was used because larger normal recurrent/hybrid weights were not practical on this machine: mlx-community/falcon-mamba-7b-4bit failed a 2K smoke with a Metal resource-limit error, and mlx-community/AI21-Jamba-Reasoning-3B-4bit required ~111s TTFT for only 2K input.

Mode	Input tokens	Output tokens	TTFT	Total time	Prompt+generation TPS	Decode TPS after TTFT	Peak memory	Profile events
Profiling disabled	100,000	10,000	10.15s	39.41s	2,790.93 tok/s	341.73 tok/s	1.65 GB	0
MLX_LM_PROFILE_RECURRENT=1	100,000	10,000	11.38s	119.69s	919.04 tok/s	92.33 tok/s	0.62 GB	241,200

Profiling-enabled run emitted 241,200 JSON records: 1,176 ssm_attn records and 240,024 metal_step records. The sum of recorded recurrent elapsed time was 105.68s. This PR does not claim an inference speedup; it provides an opt-in profiling mode, and the long run shows that default-disabled behavior avoids profiling overhead while enabled profiling captures detailed path/timing data.

Test plan

• python -m unittest discover -s tests -p test_recurrent_profile.py
• MLX_LM_PROFILE_RECURRENT=1 gated-delta smoke test
• Real-weight Mamba2 100K/10K profiling-off/on benchmark above