refactor: use time.perf_counter() for duration measurements

[m92y]

Refactor to use time.perf_counter() instead of time.time() for duration measurements.

Files modified:

• mlx_lm/cache_prompt.py
• mlx_lm/perplexity.py
• mlx_lm/quant/dwa.py
• mlx_lm/server.py

Rationale:

• Monotonicity: time.time() can be affected by system clock updates (NTP sync), whereas time.perf_counter() is a monotonic clock that never goes backward.
• Precision: Provides the highest resolution possible to calculate precise duration.
• Consistency: This aligns mlx-lm with the timing approach already adopted in the core MLX library (see [ml-explore/mlx#2943](refactor: use time.perf_counter for consistent and accurate benchmarking mlx#2943)).

Note: time.time() is intentionally preserved in server.py line 244 (self.created = int(time.time())) as this generates an epoch timestamp for API responses, not a duration measurement.