fix(ergotree-interpreter): reduce atLeast with degenerate bound instead of erroring

[mwaddip]

atLeast(bound, children) errored when bound > children.size or bound < 0. The JVM AtLeast.reduce (sigma/ast/trees.scala) reduces these to trivial props instead: bound <= 0 => TrueProp, bound > size => FalseProp (any bound, incl. > 255 — the 255 cap is a require on the child count, not the bound). The guards short-circuited before Cthreshold::reduce, which already encodes this; the reorder just lets it run, with the u8 conversion only on the valid 1 <= bound <= size path.

Found via a node syncing testnet: block 184,137 spends a V0 self-replicating contract evaluating atLeast(1, Coll[SigmaProp]()) (empty conditions) — must be FalseProp, not an error. Adds 3 regression tests.

[mwaddip]

Folded in a closely-related conformance fix on the same op: the 255-children cap.

CSigmaDslBuilder.atLeast (CSigmaDslBuilder.scala) throws MaxChildrenCountForAtLeastOp = 255 as its first statement — before AtLeast.reduce, where the degenerate-bound arms this PR adds live. So the cap overrides the degenerate reductions: atLeast(0, 256 props) must error, not short-circuit to TrueProp. ConcreteCollection carries a u16 count, so >255 children are wire-constructible and reachable.

Added input.len() > 255 ⇒ error ahead of the bound arms, with a regression test (the 256-children rejects + the 255 cap-exclusive accept). The cap is exclusive (255 accepted); the bound itself stays uncapped.