fix(transforms): apply MovingAverage multiplier to the output

[mairas]

Bug

MovingAverage's documented contract is y = multiplier * (1/n) * Σx ("Moving average will be multiplied by multiplier before it is output"). The implementation didn't honor it:

• the seeded first sample was emitted unscaled (output_ = input);
• subsequent updates applied the multiplier to the incremental delta (±multiplier * value / n), so for a steady input the multiplier cancelled entirely and only affected the transient.

multiplier = 1 (the default) behaved correctly as a plain moving average, but any multiplier ≠ 1 — e.g. the scale used in the fuel-level and milone-level examples — produced wrong output.

Fix

Maintain a running sum_ of the buffered samples and compute output_ = multiplier_ * sum_ / sample_size_ on every sample, including the seeded first one. This matches the documented formula exactly.

Tests

New test/system/test_moving_average_multiplier/:

• multiplier scales the seeded first sample;
• multiplier applies to a steady input stream (the case that previously cancelled);
• output tracks multiplier × windowed mean across several samples;
• multiplier = 1 is still a plain moving average.

Compile-verified locally with pio test --without-uploading --without-testing -e arduino_esp32 (passes). As with the rest of the suite, CI compiles tests but does not execute them; assertions were traced against the new logic and should be run on a device to confirm.

Closes #1003.

[mairas]

Automated review

Verdict: Approve. The fix correctly implements the documented contract y = multiplier * (1/n) * Σx, and all four test assertions match a hand-trace of the new logic. (Source-tracing only — CI compiles tests but does not execute them; I did not run them.)

Correctness (verified by source-tracing)

• Seeded first sample: init branch sets sum_ = sample_size_ * input, so output_ = multiplier_ * sum_ / n = multiplier * input. Previously emitted unscaled (output_ = input). Fixed.
• Steady input: sum_ += input - buf_[ptr_] leaves sum_ unchanged when input is constant, so output_ stays multiplier * mean. This is the case where the old incremental-delta form cancelled the multiplier. Fixed.
• Changing window: traced MovingAverage(2, 2.0) for set(10),set(20),set(30) → 20, 30, 50, and MovingAverage(4, 1.0) for set(4),set(8),set(8),set(8) → 4, 5, 6, 7. Both match. The steady-input test MovingAverage(4, 3.0)×6 → 15.0 also matches.
• All four assertions in test/system/test_moving_average_multiplier/ match the traced values.

State / numerics

• sum_ has in-class initializer = 0, but the init branch overwrites it unconditionally on first set(), so there is no reliance on stale state. from_json() resets initialized_ = false on a sample-size change, forcing re-seed and a fresh sum_ recompute on the next set() — no stale carry-over across reconfiguration.
• No integer-division pitfall: sum_ is float, so sum_ / sample_size_ is float division despite sample_size_ being int.
• Floating-point drift: sum_ accumulates via add/subtract deltas rather than recomputing from the buffer each call — identical drift characteristics to the previous running-output_ approach, so this is not a regression.

Compatibility

• multiplier = 1 (default): new output = sum_/n (true mean) equals the old behavior in steady state. Default behavior unchanged. Existing multiplier != 1 users (e.g. fuel_level_sensor, milone_level_sensor scale) now get correct scaling — this is the intended bug fix, not a silent behavior change for correct callers.

Conventions / minimality

• Diff is minimal: +1 member, net -2 lines in set(), plus a focused new test directory matching the existing system-test harness pattern (test_crgb, test_transform_correctness). No unrelated changes.
• The pre-existing test_transform_correctness covers from_json resize (MovingAverage::from_json resizes buffer to old size, causing out-of-bounds access #874); the new test covers the multiplier contract with no overlap.

No blocking issues. One non-blocking note already acknowledged in the PR description: assertions are source-traced, not device-executed.