Timing Context Manager

with blocks guarantee clean-up even when errors happen. Write one that measures how long a block takes, which is handy for profiling slow LLM calls or data loads.

Implement a class Timer(clock=time.perf_counter):

• Using it as with Timer() as t: records the start time on entry and the end on exit.
• t.elapsed holds the duration in seconds after the block. It's None before the block finishes.
• Timing still happens when the block raises, and the exception must propagate (don't swallow it).
• t.laps is a list. Calling t.lap() inside the block appends the time since the start (or since the previous lap).

clock is injectable so tests can use a fake clock.