fix: a gap in the input never produces a silently wrong feature value - #32

Three defects with one shape: an indicator handed a series with a gap, or evaluated
for the first time in a process, returned a finite number that was simply wrong,
with nothing raised and nothing null.

Closes #30. Closes #31.

amihud_illiquidity averaged over the whole window, not the traded bars (#31)

Amihud (2002) averages |R| / DVOL over the D periods the asset traded and divides
by D. Nulling untraded bars and taking rolling_mean(period) propagates the null
across the following window instead, so the implementation required every bar in
the window to have traded. On the NASDAQ-100 minute panel at period=30, 0.64%
untraded bars produced a 9.08% null share. The denominator now counts the bars in
the window that traded.

momentum.rsi saturated at 100 after one null (#30)

Wilder's average is recursive and had no notion of a gap. Compiled with
fastmath=True, which licenses LLVM to assume no NaN, one null returned exactly
100.0 for 87.5% of the remaining series: median 49.55 -> 100.0 on a 4,000-point
random walk. The series is now split on NaN and each gap-free run is seeded on its
own, so a gap costs the missing bar plus period warmup observations and the
oscillator resumes. On that series the gap nulls indices 500-509 and the tail
matches the clean series exactly.

fastmath was doing this across the feature surface

fastmath=True is what turned NaN propagation into silent garbage generally.
Feeding one null into every momentum, volatility and volume feature and comparing
against the clean series, 20 of 91 feature-column pairs returned a finite value
that differed from the correct one, up to 91.9 for mfi and 1.6e4 for the A/D
line. Removing fastmath leaves 6; fixing obv, ad and the polars A/D path
leaves 0. Every feature now either recovers from a gap or returns NaN.

obv froze its level forever, because both comparisons against a NaN previous
close are false. ad_polars used cum_sum, which skips nulls and carries the level
on as if the bar had been observed, while ad_numba propagates and the docstring
says it propagates. Both now propagate.

Cost, 5M rows: rsi 42.7 -> 53.8 ms, atr 51.4 -> 90.3 ms, adx unchanged. No
TA-Lib exactness test moved.

A rolling kernel compiled inside its own window

Found while investigating a downstream seal-check failure. polars.Expr.rolling_map
does not advance its window while the callback is being compiled, so a feature whose
callback calls a lazily-jitted kernel returned the first window's value for the whole
series the first time a process evaluated it. Distinct values on the first evaluation
against the second, each in a fresh interpreter with an empty Numba cache, over a
600-row series:

warm_rolling_callback runs the callback once on a probe window of the right length
so compilation happens first. Applied at all ten sites with a jitted callback,
including the four that measured clean, so no site depends on which kernels another
feature happened to compile.

Tests

• tests/features/test_missing_observation_handling.py - one missing bar against 18
  recursive and cumulative features. 23 cases fail without the fixes.
• tests/features/test_rolling_map_first_call.py - each of the twelve
  rolling_map-backed features in its own interpreter with its own NUMBA_CACHE_DIR,
  first evaluation against second. 6 fail without the warm-up.
• tests/features/test_microstructure.py - two Amihud tests, one hand-computing the
  average over the traded bars.

Full suite: 3913 passed, 1 skipped. ruff check, ruff format, ty check clean.

Note for anyone bisecting: Numba's cache=True does not invalidate when a jitted
callee changes, so delete src/**/__pycache__/*.nb[ic] after editing a kernel or
the old machine code is silently reused.