feat(fees): add computeMovingAverage for fee trend smoothing

[woahwhattheheck]

Closes #781.

fees/trend.ts exposed raw samples and percentile estimates but no smoothed
series, so callers had to implement their own smoothing to get a usable
forecasting signal.

Changes

src/fees/trend.ts — adds the exported pure function:

computeMovingAverage(samples: number[], windowSize: number): number[]

• Returns a Simple Moving Average series the same length as the input, so
  it can be plotted directly against the samples without re-aligning indices.
• The first windowSize - 1 entries are NaN. A partial average there would
  understate early movement and read as real signal.
• windowSize must be an integer >= 1; otherwise RangeError, naming the
  value received.
• Pure: the input array is not mutated.

Two implementation notes

• Non-integer window sizes are rejected, not just < 1. A fractional
  window is undefined for index arithmetic, and NaN would otherwise sail
  through every comparison and silently produce an all-NaN series. This
  matches the existing toWindowCapacity in the same file, which already
  requires Number.isInteger.
• Each window is summed directly rather than via a running total. The
  subtract-the-outgoing-value accumulator is the usual O(n) trick, but it
  drifts over a long float series. Windows here are small and bounded, so the
  exact form costs nothing worth having.

A NaN sample propagates only through the windows containing it, and the
series recovers afterwards — covered by a test.

Validation

• npx vitest run test/feeTrend.movingAverage.test.ts — 16/16 passing.
  Covers padding width, window of 1, length preservation, window larger than
  the series, empty input, input immutability, realistic stroop values, NaN
  propagation, and every rejected windowSize (0, negatives, 2.5, NaN,
  Infinity).
• npx tsc --noEmit reports 210 errors on this branch and 210 on unmodified
  main — identical, with none in src/fees/trend.ts. That baseline is
  pre-existing and unrelated to this change.