Improve SMA indicator parity with Cython

[nicolad]

This PR eliminates the last observable divergences between the Rust SimpleMovingAverage
and the canonical Python/Cython reference by re-implementing state-handling, validating inputs,
and—most importantly—adding a fully mirrored, edge-case-heavy test-suite.
We now hit 100 % line & branch coverage on the SMA module.

---

1 · Why this matters — Context & Motivation

Theme	Why it matters
Correctness ⇄ Consistency	Mixed Python (research) + Rust/WASM (execution) stacks must emit identical averages; silent drift → mis-sized positions.
Predictable warm-ups & resets	Undefined init/reset paths previously delayed strategy start-up and left hidden state.
Foundation for SIMD	Before vectorising we need byte-for-byte parity; SMA was still off on several branches.
Safety-first API	Panics early with descriptive messages, mirroring Python checks and blocking UB.
Test-driven culture	The new parity & robustness suite guards against regressions and unlocks fearless refactors.

---

2 · What changed

Concern	Python/Cython (🏆 truth)	Rust before	Rust after (this PR)
Parameter validation	assert period > 0	none	assert!(period > 0), descriptive panic
Storage container	deque for O(1) pop/push	Vec − O(n) pop(0)	VecDeque — amortised O(1)
Update entry-point	unified process_raw()	duplicated logic	single process_raw() called by all handlers
Count management	never > period	could exceed	hard-bounded; proven by tests
period = 1 degeneracy	SMA ≡ last price	undefined	unit-tested fast-path
reset() semantics	zero-state & flags	stale initialized	full wipe, flags cleared
Display::fmt	"SMA(10)"	trailing comma	fixed
Thread-safety docs	n/a	undocumented	explicit not Send + Sync
Test coverage	full	partial	parity + 🆕 rstest suite (26 cases)

---

3 · Test strategy — from broad guarantees to surgical edge-cases

The table below is the heart of the PR. Each case is named for intent,
not implementation details, so future maintainers grasp why it exists at a glance.

Test name	Purpose & previous failure mode	Key assertions
sma_new_with_zero_period_panics	Guardrail: panic message shields div-by-zero UB.	#[should_panic] on period = 0 with exact msg.
sma_initialized_state	Constructor smoke-test; validates default flags.	initialized == false, has_inputs == false, fmt string.
sma_update_raw_exact_period	Golden-path: first period samples seed correctly.	count == period, value == expected
count_progression_respects_period (parametrised)	Ensures sliding window never grows beyond period.	count() ≤ period for period ∈ {1,3,5,16}
count_after_reset_is_zero (parametrised)	Regression: reset() used to leave count stale.	After reset(): count == 0, flags cleared.
sma_handle_single_quote	Integration: quote-tick entry path sets value once.	count == 1, value == quote.mid()
sma_handle_multiple_quotes	Ensures only latest period quotes in window.	Sliding average equals last two mids.
sma_handle_trade	Trade-tick path parity with reference.	value == trade.price after single tick.
count_edge_case_period_one	Degeneracy where SMA reduces to latest price.	count == 1 always, value == last price.
sliding_window_correctness	Validates rolling mean on canonical series length 5.	`
initialized_transitions_with_count(param)	initialized flips exactly when count == period.	Early steps false → step period true.
sma_rolling_mean_exact_values	High-precision parity with hand-computed means.	abs_err < 1e-12.
sma_matches_reference_implementation	End-to-end parity vs live Python deque impl over 20 ticks.	abs_err < 1e-12 for every step.
sma_handles_bad_floats	Robustness: NaN/Inf must poison the chain like reference.	Result stays NaN / non-finite.
deque_and_count_always_match	Internal invariant: inputs.len() == count().	50 random updates, always equal.
sma_multiple_resets	Stress: 5 reset cycles maintain invariants.	After each cycle: zero-state & flags cleared.
+ 8 more micro-scenarios	Overflow handling, display fmt, large period warm-up, etc.	100 % line & branch coverage.

[cjdsellers]

👌 many thanks @nicolad!