Fix #233: detect plot overlay period from data frequency instead of timestamp precision

[Geochelone666]

Fixes #233.

Problem: daily bars whose last timestamp carries seconds (e.g. non-midnight
timestamps) were misdetected as intraday data, so the automatic OHLC overlay in
Backtest.plot() raised instead of superimposing.

Fix: pick the resample rule from the existing _data_period helper (median
of the last 100 datetime deltas) using period thresholds (day -> ME, hour -> D,
minute -> h, second -> min, millisecond -> s), instead of guessing the sampling
period from timestamp precision. Explicit string rules, the upsampling
ValueError, and the same-period skip behavior are unchanged.

Tests: 3 new regression tests in backtesting/test/_test.py (daily data with
seconds on the last bar plots fine; explicit hourly upsampling still rejected;
OHLC aggregation of the superimposed Bokeh source verified across 5 real periods
with microsecond offsets). Full TestPlot suite passes on master baseline
ca2e261; flake8 clean; mypy output identical to baseline.

Known limitations: the median only looks at the last 100 points, so
irregular/duplicate timestamps aren't guaranteed a unique true period
(pre-existing helper limitation).

AI disclosure: this patch was drafted with AI assistance (OpenAI Codex) and
reviewed by a human before submission.