Improve EMA indicator parity with Cython

[nicolad]

Improve EMA indicator parity with Cython

Restores feature-parity and behavioural equivalence between the Rust ExponentialMovingAverage (EMA) implementation and the canonical Python/Cython reference.

---

1 Why this matters — Context & Motivation

Theme	Why it matters
Correctness ⇄ Consistency	Mixed Python (quant research) + Rust/WASM (execution) stacks must produce identical indicator values or silent PnL drift creeps in.
Predictable warm-ups & resets	Edge-cases (e.g., period = 1, reset() after NaN) previously left the indicator in an undefined state.
Baseline for SIMD hot-paths	Before vectorising we need a byte-for-byte port; this PR closes the last gaps.

---

2 What changed

Concern	Python/Cython (🏆 truth)	Rust before	Rust after (this PR)
Parameter validation	✔ period > 0	✘ none	✔ assert!(period > 0)
alpha for period = 1	✔ α = 1 → immediate tracking	✘ α < 1	✔ α = 1, unit-tested
First-sample handling	✔ initialises & count = 1	✘ count = 0	✔ aligned
count increments	✔ every update	✘ skipped on first	✔ parity
reset() semantics	✔ full zero-state	✘ stale has_inputs	✔ cleared
Display::fmt	n/a	✘ extra comma	✔ fixed
Thread-safety docs	n/a	✘ missing	✔ explicit not Send + Sync
Test coverage	full	partial	full parity + 🆕 rstest suite

---

3 Five practical ways to put EMA to work

#	Use-case	Why EMA helps	Wiring sketch
1	Classic cross-over	Baseline lag is acceptable for many trend followers.	long when price > EMA, exit on cross-down.
2	Dynamic position sizing	EMA slope ≈ momentum; scale size by slope / ATR.	size = k · d(EMA)/dt.
3	Mid-price estimator in MM engines	Smooths quote flicker better than mid-ticks.	quote_mid = EMA(last_trade).
4	Volatility gating	σ(price – EMA) ≈ realised σ without extra filters.	Open only when σ < threshold.
5	Time-series momentum screener	Rank assets by close / EMA – 1.	Nightly batch pipeline.

---

4 Implementation notes

• Guardrails — period > 0 panic matches Python.
• Period = 1 fast-path — α = 1, so EMA == last sample; avoids divide-by-zero.
• has_inputs lifecycle — toggled on first sample, cleared on reset().
• NaN poisoning — once a NaN enters, value stays NaN until reset().
• Thread-safety — indicator itself is not Send + Sync; wrap in Mutex/Arc for multi-threaded use.
• Public API — no breaking signature changes; only stricter panics.

---

5 Tests added / updated

• test_period_one_behaviour — α = 1 behaviour
• test_nan_poisoning_and_reset_recovery
• test_has_inputs_lifecycle
• test_subnormal_inputs_do_not_underflow
• Regression: new_panics_on_zero_period

---

Related: #2507

[cjdsellers]

Great work @nicolad, many thanks!