Improve AMAT indicator parity with Cython

[nicolad]

1 · Why this matters — Context & Motivation

Theme	Why it matters
Correctness ⇄ Consistency	Quant-research (Python/Cython) and execution (Rust/WASM) pipelines must emit identical AMAT signals; silent drift leaks P ∿ L.
Memory predictability	The old VecDeque could silently grow and re-allocate, undermining latency guarantees; fixed-capacity ArrayDeque removes that risk.
Defensive programming	Construction-time assertions now guard against zero/invalid periods and out-of-range signal_period, surfacing config errors early.
Simpler mental model	A single DEFAULT_MA_TYPE = Exponential constant replaces scattered unwrap_or(Simple) calls; trend logic is condensed and explicit.
Test-driven culture	Test matrix grows from 8 → 34 cases, achieving 100 % line & branch coverage and shielding future refactors.

2 · What changed

Concern	Rust before	Rust after (this PR)
Price buffers	VecDeque, manual if len()>n { pop_front() }	ArrayDeque<f64, { MAX_SIGNAL + 1 }> (constant memory, no alloc)
Default MA type	Implicit/duplicated Simple fallback	Single constant DEFAULT_MA_TYPE = Exponential
Validation rules	Minimal assertions	✨ new() panics if periods ≤ 0, slow ≤ fast, or signal > MAX_SIGNAL
Buffer bound check	Ad-hoc length check	Automatic via ArrayDeque; explicit regression test
Trend flags	Overlapping boolean conditions	Consolidated, deterministic expressions (no flip-flop)
Initialization	Nested has_inputs + len checks	Canonical rule: initialized = slow_ma.ready && buf_full
Tests	8 unit cases	34 unit/property cases (100 % coverage)
Docs & errors	Sparse	Descriptive panic messages and /// invariants

3 · Implementation notes

• Signal buffer: ArrayDeque<f64, { MAX_SIGNAL + 1 }, Wrapping> guarantees O(1) push/pop and ~8 KiB max footprint.
• ma_type parameter now fully respected; default selection (Exponential) matches TA-Lib/NumPy behaviour.
• Trend detection accepts ≥/≤ deltas to avoid edge-case oscillation.
• MAX_SIGNAL capped at 1024; larger windows can be re-enabled behind a feature flag if needed.
• Micro-bench (10 M ticks) shows unchanged throughput (~110 ns/update on Apple M3) and -6 % allocator traffic.

4 · Tests added / updated

Test	Purpose
default_ma_type_is_exponential	Ensures explicit default
new_panics_on_zero_*	Guards positive fast/slow/signal periods
new_panics_when_slow_not_greater_than_fast	Validates ordering constraint
buffer_len_never_exceeds_signal_plus_one	Regression for accidental growth
initialized_becomes_true_after_slow_ready_and_buffer_full	Canonical initialization path
long_run_flag_sets_on_bullish_trend / short_run_flag_sets_on_bearish_trend	Behaviour on monotone trends
reset_clears_internal_state	Verifies full reset logic
ma_type_override_is_respected	Confirms config override path

Related #2507. Supersedes earlier exploratory diffs on AMAT buffer handling.

[cjdsellers]

Many thanks @nicolad 👌