bench: sample turns in microseconds instead of whole milliseconds

[mparrett]

xsofy.perf/time-turn timed each turn with System/currentTimeMillis, so every per-turn sample was floor-quantized to a whole millisecond before any statistic ran. On a ~9 ms turn that hides anything smaller than ~10% from every downstream percentile — and the quantization error is correlated, so summing samples doesn't recover it. We hit this A/B-ing lg builds against the pinned release: two builds whose real difference was ~0.8 ms/turn read as identical 9 ms columns.

Changes:

• time-turn samples with System/nanoTime and returns integer microseconds. Integer µs keeps aggregation exact, stays readable in the dev console's 'bench dump, and fits a 32-bit int (~35 min/turn) where nanoseconds would overflow at ~2.1 s on TinyGo-class targets.
• summarize maps carry :unit :us so consumers don't have to guess the unit from magnitude.
• bench/native.lg prints ms with three decimals (avg-ms/p50-ms/... headers instead of a per-cell suffix).

Same build, before and after:

:wait           285     9ms     9ms    12ms    22ms      (before)
:wait           285     9.966     9.252    13.914    40.708  (after)

Verified: full test suite green (318 tests, 0 fail); the bench runs under both the pinned v1.12.2 lg and a tip build; the console 'bench command still returns in well under a second.