perf(engine): expanding-mean ADV fallback during intraday warm-up (follow-up to #269/#272)

[shardul0701]

Follow-up to #269/#272. Raised as an open question in #272's review: min_periods was deliberately left at 1 because raising it to a full session would leave the first session of every intraday backtest with no liquidity cap and no market impact at all (both _daily_adv consumers guard with if pd.notna(adv_20) and adv_20 > 0, so NaN skips the cap block rather than applying a cautious one).

Problem

During warm-up (fewer than _adv_bars_per_day * 20 bars of history), _daily_adv() extrapolates from whatever partial window it has, weighted by min_periods=1 — in the worst case, a single bar's volume × bars-per-day. That's a fair, unbiased but high-variance guess, and it's strictly better than the pre-#265 behaviour, but it's not the best available estimate.

Proposed fix

Use an expanding mean during warm-up instead of a fixed-window rolling mean with min_periods=1:

• For at_date where fewer than _adv_window_bars are available, average all bars seen so far (rescaled to daily by _adv_bars_per_day) instead of relying on min_periods=1's partial-window behaviour.
• Once _adv_window_bars of history exists, behaviour is identical to today — this only changes the warm-up period.
• Must compose with perf(engine): memoize the per-symbol ADV series #272's per-symbol memoization (the cached series should already reflect the expanding-then-fixed window; do not add a second cache layer).
• Does not change trade admission logic — no trade that currently fires should stop firing, and no trade that's currently blocked should start firing. Only the cap/impact magnitude during warm-up changes.

No-regression requirement

• D/W/M timeframes: bpd == 1, so the "warm-up" period is the first ~20 calendar days of any run — same as today's behaviour for daily backtests, since 20 daily bars is a short warm-up already covered by existing tests. Confirm no numeric drift there.
• Any backtest whose full history already exceeds 20 days behaves byte-identically (the expanding-mean path is only reachable during the literal first _adv_window_bars bars of a series).
• Golden master (tests/test_engine_characterization.py) green.

Acceptance criteria

• Expanding-mean fallback implemented for the warm-up period, composed with perf(engine): memoize the per-symbol ADV series #272's per-symbol cache
• Test proving warm-up-period cap/impact values match a hand-computed expanding mean, not a single-bar extrapolation
• Test proving trade admission (which trades fire) is unchanged versus the pre-fix behaviour on a fixture with a real warm-up gap
• Test proving post-warm-up behaviour is byte-identical to today
• Golden master green; full suite green