Close the remaining native Array performance gaps (#505)

[ciaranra]

What

Fixes the remaining #505 items at the layer that owns each cost: fancy (integer-list) reads, boolean mask reads, array(list) construction, and astype widening. flatten() and f64->f32 were profiled, found already at NumPy speed on this host, and left untouched with tightened guards.

Performance (release build, 1e6 elements, best-of-5, PECOS/NumPy ratio)

operation	before	after
fancy read (every-7th list)	11x	0.52x
boolean mask read (~50%)	14x	1.30x
array(int_list)	14x	0.37x
array(float_list)	32x	0.35x
astype u8->i64	2.0x	1.07x
flatten() 1000x1000	1.06x	unchanged (already copy-bound: 31% memmove)

Root causes (per-item perf profiles in the implementation report): per-element Python protocol dispatch during index/list parsing, per-element coordinate-vector allocation in selection, and a redundant full validation scan before provably-safe widening casts. Fixes: single-pass native index normalization, logical C-order offset gathering, native mask compaction, preallocated typed buffers for homogeneous built-in lists, and a range-subset bypass (source integer range contained in target integer range, computed in i128) that skips validation only when overflow is impossible -- narrowing still validates (300 -> uint8 still raises naming the value).

Guard ceilings tightened, never loosened

test_performance_guard.py ceilings drop to ~10x the new medians (fancy 200ms->20ms, mask 300ms->50ms), with new guards for both list constructors and f64->f32; 15/15 pass on release.

Verification (independent of the implementation run)

• Full pecos-rslib suite: 2,103 pass; the 63-config bit-exact decode-equivalence fixture suite passes on the rewritten indexing paths.
• Adversarial randomized fuzz vs the NumPy oracle: 2,700 comparisons over fancy reads (negative/repeated/empty/out-of-range), masks (densities, wrong-length raises, 2D Fortran layout), and casts -- zero divergence.
• List-construction dtype inference spot-checked against NumPy: homogeneous lists all match; the mixed-type divergence found during review ([True, 1] -> bool) reproduces identically on dev and is filed as Mixed-type list construction infers dtype from the first element instead of promoting like NumPy #539, not introduced here.
• Cold cargo clippy -p pecos-rslib --all-targets -- -D warnings, cargo fmt --check, unshielded pre-commit run --all-files: clean.

Implementation by Codex from a profile-first task packet; root causes, semantics, and benchmarks independently verified.